What's New

Saving chats now uses far less memory and large imports run faster. LLMnesia also releases its search model's memory when you stop searching and takes about 10 MB less space once installed. The first time this version starts, it reorganises your search index in the background; search keeps working while it does.

• Google AI Studio chats save automatically when opened, and older chats can be imported straight from AI Studio or from files downloaded from Google Drive.
• Character.AI imports include every chat with a character, archived chats, and the alternative replies on each chat's latest message.
• Import history starts straight away, and new chats from your other devices arrive on their own more often.
• Search results no longer flash after they appear, very large Claude Code and Codex sessions and the automatic backup use far less memory, and Settings has a Copy logs button for bug reports.

Google AI Mode imports now continue through the Google tab when background access is unavailable, preserving saved progress and protecting unsent messages. Completed replies save their later paragraphs and lists, original-chat links reopen the right conversation, and the save indicator settles correctly. Keep the Google tab open while importing; LLMnesia cannot recover text Google does not load.

• Clearer import counts, Resume import and individual chat retries make interrupted imports easier to finish.
• Outdated import, memory-building and save warnings clear when the underlying problem has ended.
• Vault uploads and deletions remain recoverable through overlapping sync attempts and browser restarts. Older local-session renames preserve newer messages.
• Saved copies refresh while preserving selected text, Google reply emphasis and Gemini formulas. This update also improves chat-title updates, Vault feedback and optional first-search suggestions.

Large Google AI Mode imports recover more reliably when older history pages load slowly. Imports started from the empty AI Mode page stay on the final conversation instead of returning to a blank page. Vault status shows clearer activity and counts, explains skipped items, and clears successful sync feedback automatically.

Google AI Mode history imports continue past chats that fail to open, leaving them available for retry. Claude no longer repeats capture feedback when nothing has changed.

Reliability improvements to Vault sync and conversation indexing.

Import Google AI Mode conversations from your history, including older chats beyond the first ten. Interrupted imports resume unfinished work, slow-loading chats have a longer individual retry, and saved conversations preserve visible replies and formatting with fewer duplicates. Vault sync and restore are more reliable, with responsive status details in the popup and Settings. This update also improves sign-in and reduces repetitive background analytics.

• Import your Google AI Mode history from its sidebar. Enable Google AI Mode in Settings first.
• Fix missing DeepSeek messages and delayed Copilot replies during imports.
• Avoid duplicate Google chats and false warnings about stalled imports.
• Wait for the correct conversation before saving, and flag incomplete imports.
• Save images that load after the message text.

Copilot history imports now wait for the correct chat and for both questions and replies to load, preventing messages from being saved under the wrong conversation or missing delayed answers. Imports also handle already-open chats and restore your original chat more reliably. If a chat is only partly captured, LLMnesia preserves its fuller saved copy and keeps the warning visible when the import finishes.

Source links from your desktop AI answers now discover the installed LLMnesia extension instead of trusting the ID saved in the link, so they keep working after a reinstall or a move between browsers. From a saved ZCode chat you can open its project folder, or copy a search phrase that finds the chat inside ZCode. Copilot support covers its new domain and message layout, including history import for the conversations it discovers, and search now matches text in languages whose words are not separated by spaces. Vault errors show support codes for easier troubleshooting, large exports split between whole conversations instead of through the middle of one, and this release also hardens Vault passphrase changes and sync against interrupted or out-of-date devices, keeps partially captured chats marked for retry, and makes ChatGPT Continue-in and local coding-session imports more dependable.

You can now turn on daily backups in Settings to keep extra copies of your conversations in this browser profile. History imports start more reliably from the popup and onboarding, and explain when browser site access blocks them. If a chat cannot be saved, LLMnesia keeps your existing conversations and shows a persistent warning; deleting local data now requires typing DELETE. You can also show a Vault passphrase while entering it, and Codex replies are clearer in saved transcripts.

Your ZCode prompts and assistant replies are now indexed on this device and searchable alongside Claude Code and Codex sessions. Ask Vault picks up conversation changes about 30 seconds after they settle. This release also removes markup codes from saved text, syncs ChatGPT Project moves across devices without resending the conversation, and makes your messages easier to distinguish from the assistant's in ZCode and Perplexity Vault chats.

A change to how ChatGPT displays messages interrupted conversation capture. LLMnesia now saves new ChatGPT messages to your searchable history again.

If you forget your Vault passphrase, use the recovery key you saved during setup to choose a new one and unlock the same encrypted archive. Recovery accepts keys pasted from notes, verifies that the key belongs to your Vault before changing anything, and works while conversations are still syncing.

You can now sign in to the Vault web app with a passkey: Face ID, Touch ID, or Windows Hello can replace the emailed code on devices that support them, and a passkey only proves who you are, so your passphrase and encryption are never touched. Subscribing to Vault no longer leaves the extension either: Settings shows the real prices, and checkout and “Manage subscription” happen on Stripe’s own pages. The web app is also more dependable day to day. You stay signed in between visits, it picks up conversations saved since your last visit, and interrupted downloads carry on from where they stopped. Reliability fixes round it out. A Perplexity import no longer quietly restarts in the background all day, an import that stops early now says how far it got instead of reporting itself finished, and the desktop connector keeps the same searchable history in a much smaller format.

Images in your ChatGPT conversations are now saved along with the chat, so your saved copies and search show the whole conversation as you saw it. Signing in to Vault now makes clear you’re opening your existing Vault — not creating a second one — and shows syncing working as it happens. Vault is also more dependable: conversations that fail to reach the sync queue are retried, missed chats are picked up automatically, and a full restore clears out-of-date warnings about conversations that couldn’t be read. LLMnesia’s desktop assistant connection now also supports ZCode alongside Claude Desktop, Claude Code, Cursor, and Codex.

Meta AI chats can now be saved and searched alongside your other AI history, including an import of conversations visible in Meta AI’s sidebar. You can also give saved conversations private, searchable titles without changing the original chats. Google AI Mode access is now opt-in, and Gemini capture is more reliable during long, updating answers.

On supported devices, you can choose to unlock Vault with your device biometrics instead of entering your private passphrase each time. You can also save full-page Google AI Mode conversations alongside the rest of your history.

A small reliability release: Character.AI history imports now work with the site’s latest changes, use the same “Index older chats” prompt as other supported platforms, and explain clearly when a retry is needed. Desktop AI source links also now always open a conversation, using LLMnesia’s local Viewer when the original chat no longer has a usable link.

Your local Codex sessions — from the ChatGPT desktop app, Codex CLI, and Codex IDE extension — are now indexed and searchable right alongside Claude Code, capturing only your prompts and Codex's replies, never file reads, commands, or tool output. This release also fixes a Grok import that could quietly stop after ~60 chats and report itself complete, keeps search fast as your index grows, makes history imports more resilient across browser restarts, and fixes long Qwen chats splitting into duplicate conversations when scrolling revealed older turns.

Importing your ChatGPT history reads roughly three times faster, and no longer grinds for hours when ChatGPT rate-limits it. When an import does pause, it no longer looks like something's broken — it quietly resumes on its own, with a one-time nudge toward exporting your history only if that genuinely doesn't clear up.

You can now copy just one message out of a saved chat instead of the whole transcript, and the popup gives a quiet nudge when one of your platforms hasn't been imported yet — plus fixes to ChatGPT import reliability, search suggestions, and starred-chat search.

Search is noticeably faster on large libraries, plus a batch of indexing and reliability fixes across Qwen, Gemini, and Settings.

• Highlight text in any chat to save it as a memory. Select a passage in a conversation and a small "Save to Memory" button appears. Give it a label, save, and it joins your memories, ready to drop into a future chat. Each saved passage remembers where it came from, so you can jump straight back to that point in the original conversation. This works on every site LLMnesia supports, not just a couple of them.
• Search straight from the toolbar popup. Opening the popup now shows a search box right away, so you can find a past chat without needing to be on a supported site first. Type a few words, and results show which platform each chat is from
• The toolbar icon now shows how many chats are indexed. LLMnesia works silently in the background, and it used to be easy to install it, close the welcome tab, and never realise anything was happening. Once you've got a handful of conversations saved, the extension's toolbar badge now counts them live — "5", "47" — so the invisible work becomes a visible, climbing number. It disappears the first time you actually use it (open the popup, run a search, or drop in a memory). If clicking the icon is what made it disappear, the popup explains what that number was, right above the search box.
• Give a memory its own short label when you create it. The add-memory form now has an optional label field, so you can name a memory (e.g. "Code style") instead of it always taking the first few words of the text. Leave it blank and the label is generated from the memory as before.
• Z.ai is now supported. Your chats on chat.z.ai are saved and searchable alongside the rest, with both sides of the conversation kept in order and just the final answer indexed (not the collapsed "Thought Process" reasoning trace).
• One-click import for your Z.ai history. As well as saving new chats as you have them, LLMnesia can bring in your existing Z.ai conversations in one go, the same way it does for ChatGPT, Claude, Gemini, Perplexity, Grok, and Character.AI. Open chat.z.ai and start the import from the welcome screen or the toolbar popup.
• One-click import for your DeepSeek history. As well as saving new chats as you have them, LLMnesia can now bring in your existing DeepSeek conversations in one go, the same way it does for ChatGPT, Claude, Gemini, Perplexity, Grok, Character.AI, and Z.ai. Open chat.deepseek.com and start the import from the welcome screen or the toolbar popup.
• Drop a memory into a chat without clicking. Hold Option (Alt) for a moment on any chat and a numbered list of your memories appears by the composer. Let go, press that number, and the memory is injected straight into your message. No picker, no panel, no clicks. Your own typing is untouched: the number is pressed on its own rather than with Option held, so Option and a number still produce the character it always did on your keyboard (a "#" on a UK Mac, an "@" on a Spanish one). Hold Option again to hide the list without picking anything. The picker itself also has a "Clear all" button to reset your selection in one go, and each memory now has a delete button right next to its edit pencil — one click arms it, a second confirms, no trip to Settings needed.
• When an import finishes, LLMnesia now offers you the search straight away. Bringing in your history is the slow part, and the moment it finished you got nothing back: the "try a search, N chats now indexed" nudge waited a full day before it first appeared, and only then if you happened to be back on one of your chat sites. So the one moment your whole history became searchable passed in silence. That nudge now appears the moment the import finishes, in the tab that just did it, with the reminders that follow moved to a day and three days later.
• Search now offers to bring in your history, instead of telling you where to look. Opening search before anything is indexed used to show a line explaining that you should open a chat tab and find "Index older chats" in the extension popup. If you hadn't pinned LLMnesia, that was a dead end. There's now a button right there in the search box: on a supported site it starts importing that site's history on the spot, and anywhere else it offers to open ChatGPT or Claude and start there. If an import is already running, it says so rather than offering to start another.
• A nudge to pin LLMnesia to your toolbar, only if you haven't. The welcome tab now shows a short walkthrough of where the pin control lives (Chrome's puzzle-piece menu), and confirms the moment you do it. If the icon is already pinned, nothing appears. There's also an occasional one-line reminder at the bottom of search, which you can dismiss for good.
• The popup shows you one thing at a time. The popup has a number of cards that each appear only when they're relevant: a setup checklist, a "what's new" note after an update, a prompt to import your older chats, and so on. Each decided for itself whether to appear, so on an unlucky open you could get five or six of them stacked above everything else. Now at most two can show together, one note and one suggestion, picked in a sensible order, and the rest wait for another day. Nothing is lost: a card that stands aside comes back on the next open.
• The toolbar popup is calmer, and opens on your recent chats. The space under the search box used to sit empty until you typed something. It now lists your four most recent conversations the moment the popup opens, so you can jump back into what you were just doing in one click, and clearing the box brings them back. The rest of the popup lost a lot of clutter along the way: the keyboard shortcut was being explained in three different places and is now shown once, "Indexer Status" and "Index older chats" have become a single Status panel, the empty progress bar only appears once there's actual progress in it, the "Max chats to index" box is tucked under Options, and "Report a problem" and "Visit site" have moved out of the header down to the footer next to Full settings.
• Snippets and Memory are now one feature. Snippets was an early beta for saving passages out of a chat, but a saved snippet just sat in a list of its own. Saving a highlight now creates a memory instead, which means it can be added to future chats, gets included in your backups, and lives in one place with everything else you've saved. Any snippets you already had are moved across automatically the first time you open the extension, keeping their labels and their link back to the original chat. They arrive switched on but not automatic, so nothing new starts being added to your chats without you choosing it. The separate Snippets section in Settings and the Snippets tab in search are gone.
• Claude chats where you pasted in a block of text could lose track of who said what. When a Claude turn carried pasted content or an attachment alongside your message, LLMnesia had to save the whole turn rather than the message on its own, and those turns arrived with no speaker attached. Both sides of the conversation were still saved, so nothing looked obviously wrong, but the affected turns were stored unlabelled, which meant "Continue in" could hand the wrong half to the next assistant as if you had said it. Speakers are now worked out from the turn's contents when the label is missing, and a conversation that comes back with a lot of unlabelled turns is reported the same way a conversation missing a whole speaker already was.
• Character.AI and Grok were called by the wrong name in a few places. The import prompt and its status messages kept their own list of platform names, and that list had not kept up: on Character.AI and Grok, messages like "Some recent Perplexity chats aren't searchable on this profile yet" or "Queued Perplexity backfill" named a platform you weren't using. Settings had the same problem and called Character.AI "Grok". Every one of these now reads from a single list, so a newly supported platform cannot show up under an older platform's name.
• Long Qwen conversations could quietly lose their earlier messages. Qwen only loads recent messages into the page at first, and reveals older ones as you scroll up. On a long chat, LLMnesia could catch the page in that partly-loaded state and mistake it for the whole conversation, overwriting the real earlier messages with nothing. It now recognizes a partial view of a chat it already has and leaves the saved copy alone until it sees the full conversation again.
• A very long single message (e.g. a big pasted block of context) could get saved cut off partway through. A message that long can take a moment to finish rendering on the page, and if LLMnesia captured it in that in-between state, the incomplete version stuck permanently, since nothing prompted it to try again. It now recognizes when a freshly captured message is just an incomplete read of a longer one it already has saved, and keeps the fuller version instead of overwriting it with the partial one.
• Starring a chat could appear to work but not actually save. If the chat wasn't found in your index at the moment you tapped the star (for example, it had just been removed from view), the star would show anyway and then vanish the next time you opened the search. Now a star that can't be saved is rolled back right away with a clear "Star failed" message, so you always know whether it stuck.
• Qwen chats could be saved under the wrong title, or with "Thought completed" instead of the actual reply. On some Qwen pages the browser tab's title is just the generic product name rather than anything about your conversation, so that generic name was being used as the saved title instead of falling back to your first message like it does elsewhere. Separately, a reply that showed Qwen's "thinking" status card was sometimes saved as just that status label ("Thought completed") rather than the answer that followed it. Both are fixed: the real title or reply text is now captured correctly.
• Picking multiple saved memories to add to a chat could silently leave some out. In the "Add my context" picker, anything past an internal size limit was dropped with no warning, so checking several memories and adding them could quietly add fewer than you picked. Every memory you select by hand is now included, and a brief confirmation shows how many were added.
• A memory's On / Off / Auto setting in Settings wasn't reflected in the memory pill. Once you had interacted with the pill on a platform, its saved selection overrode the Auto flag, so switching a memory to Auto in Settings didn't pre-select it in the pill, and the chip list gave no visual hint either way. Auto memories are now always included in the selection regardless of prior picks, and each chip in the pill shows a small "Auto" badge so the setting is visible at a glance.

• Character.AI is now supported. Your chats on character.ai are saved and searchable alongside the rest, so a line a character said (or something you told them) turns up in the same search as your ChatGPT and Claude conversations. It keeps both sides of each chat with the right speaker, and you can filter search to just your Character.AI chats.
• One-click import for your Character.AI history. As well as saving new chats as you open them, LLMnesia can now bring in your existing Character.AI conversations in one go, the same way it does for ChatGPT, Claude, Gemini, Perplexity and Grok. Open character.ai and start the import from the welcome screen or the toolbar popup, and your past chats become searchable without you having to reopen each one. It brings in your active chat with each character (the one that opens from your recent chats), keeps the whole back-and-forth in order, and skips anything already saved so nothing is duplicated.
• Search your Claude Code sessions too — Settings now has an "Index Claude Code sessions" option that adds your local Claude Code chats — from the terminal, the VS Code extension, and the desktop app — to the same search box as your ChatGPT, Claude, and Gemini conversations. Point it at your ~/.claude/projects folder and it pulls them in. For a one-click, no-folder-picking setup, an optional helper app lets LLMnesia keep your sessions in sync on its own in the background: start a new Claude Code chat and it turns up in search without you doing anything. Only your prompts and Claude's replies are indexed — file reads, commands, and tool output are left out.
• Pick a Claude Code session back up where you left off — a saved Claude Code chat has no web page to reopen (it lives on your device), so its viewer now offers two ways back in: one click reopens it in the Claude Desktop app, or copy a ready-to-run terminal command that resumes that exact session in the right project folder.
• Claude Code indexing is easier to find — connecting your Claude Code sessions used to be tucked away in Settings. Now the extension popup has a one-click "Connect your Claude Code sessions" shortcut that takes you straight to setup, and if you search for something code-related and it isn't there yet, the search box gently offers to add them. Both quietly step aside once you've connected, and you can dismiss them if Claude Code isn't for you.
• Import a large ChatGPT history from your export file. On a big ChatGPT account, the automatic import can be slow and may not reach everything, because ChatGPT's site limits how fast it hands over older chats. You can now bring in your whole history in one go: export your data from ChatGPT (its Settings → Data controls → Export data), then load the .zip it emails you into LLMnesia's Settings. It reads only your chat text (images and other files in the export are ignored), updates chats you've already imported in place rather than duplicating them, and picks up even archived chats the live import can't reach. For a large history this is the complete, rate-limit-proof route. A progress bar shows how far you are, and a Stop button lets you pause and pick up where you left off next time (the same .zip will skip the chats you've already saved and finish much faster). When the automatic import keeps hitting rate limits on a very large history, the corner pill now points you straight to the export importer as a faster path.
• Backup export and import now show a progress bar, and a stopped import picks up where it left off. In Settings → Backup & Restore, exporting or restoring a large index used to give little more than a "working on it" message with no sense of how far along it was. Both now show a progress bar. Import also gets a Stop button, and if you stop it partway through (or the tab closes before it finishes), choosing the same backup file again skips everything already restored and continues from there instead of starting over.
• Restoring a large backup (or any bulk import) is now much faster. Saving each conversation used to check and update your search index one small piece at a time, in sequence — for a chat with hundreds of words that added up to a lot of waiting, and on a large backup it could take well over an hour. Those checks now happen together in batches instead, with no change to what gets saved.
• Importing older chats now stays quietly in the background. Bringing in your history is long-running work that pauses to wait out the AI's rate limits, and it used to hold the little corner indicator wide open the whole time: first a running "backfilling 214 of 1962" count, then a "rate limited, resuming in 45s" line ticking down second by second. For something meant to just tick along in the background, that read as loud and a little alarming. It now settles back to its normal resting dot, with a soft pulse to show it's quietly working, and you can hover it any time to see how far it's got. It still speaks up briefly for the things worth a glance, like a "5 chats indexed" when a stretch finishes, or a sign-in prompt if it needs one.
• Your saved chats are kept as your library grows, instead of being trimmed to save space. To stay under a storage limit, LLMnesia used to remove your oldest chats (the ones you hadn't starred) once your saved library passed a certain size, quietly dropping them from search to make room. That limit is now far higher, with room for many times more of your history, and LLMnesia now asks your browser to keep your saved chats durably rather than treat them as temporary data it might clear when disk space runs low. In practice your full history stays searchable, and starred chats remain protected as before. The manual storage-size limit that used to sit in Settings has been retired, since it no longer serves a purpose now that your chats are kept safely.
• Importing a long history no longer stops partway through and reports success. LLMnesia reads your conversation list a page at a time, and a big history takes well over a hundred of those requests. The further back in your history a page sits, the longer the AI's site takes to hand it over: on a 3,600-chat ChatGPT account the earliest pages took over three times as long as the most recent ones. LLMnesia was giving up on any page that took more than five and a half seconds, so on a large account the import stopped somewhere in the middle of your history, kept only the chats it had reached, and still told you it had finished. Older chats were simply missing from search, with nothing to suggest anything had gone wrong. LLMnesia now waits properly for those slower pages and allows more time on each retry, so even very large histories are read to the end. It also retries a failed request instead of abandoning the whole import, and when it genuinely can't finish reading your list it records the import as incomplete rather than claiming it worked. This applies to ChatGPT, Claude, Gemini, Perplexity and Grok.
• A rate limit during an import no longer ends it. Reading a long history means asking the AI's site for a lot of pages, and part-way through it can tell LLMnesia to slow down for a while. LLMnesia treated that as a normal error: it tried again four times over about three seconds, which is nowhere near long enough for the site to start answering again, then stopped and reported the import as finished. Whatever it had reached by that point was all you got, so re-running the import could leave you with a different, arbitrary slice of your history each time. LLMnesia now recognises being asked to slow down, waits it out properly (up to a minute and a half per page, and it obeys the site's own requested wait when it sends one), and carries on where it left off. It also stops making four requests where one will do, which is what pushed it into the limit sooner than necessary. If the site is still holding it off after all that, the import is recorded as incomplete and blamed on the rate limit, so it's clear that simply trying again later will work. The same applies while it's saving each chat: being asked to slow down there used to count as three chats failing in a row, which LLMnesia read as "something is badly wrong" and used to abandon the entire import, sometimes with thousands of chats still to go. Two related gaps are fixed alongside it: a Claude import that hit a snag while listing your workspaces reported an empty account as though it had finished, and a Gemini import could quietly drop all of your pinned chats while reporting success.
• An incomplete import no longer counts as a finished one. LLMnesia marked a platform as imported as soon as a run saved a single chat, so an import that was cut short part-way through quietly retired itself: it never offered to import your history again, and there was nothing anywhere to say it had only got a fraction of it. Combined with the problems above, that's how a large history could end up permanently missing its older years while everything looked normal. A platform is now only marked as imported once a run has actually reached the end of your history. If it stopped early, LLMnesia will offer to pick it up again.
• A large ChatGPT import now drives itself to the finish, with no babysitting. Reading a very long history can take a while, and a rate limit will pause it part-way. Until now every attempt started over from the top of your list, so a pause meant re-reading all the same pages before reaching new ground (spending the very quota that tends to cut these imports short), and picking it back up was on you: you had to remember to start it again. Now LLMnesia keeps its place and continues from where it stopped. When a rate limit pauses a run, it waits the limit out and starts the next stretch on its own, right there on the open tab, without taking over the corner indicator to do it (it stays a quiet, softly pulsing dot you can hover for progress), and if you close the tab it still picks up on your next visit. It also imports more gently now (one chat at a time, well spaced, with longer waits between stretches), so it provokes the rate limit far less often in the first place. Because it reads your newest chats first, the ones you're most likely to search for are in place after the first stretch while the older years fill in over the following ones. Two things that used to derail this are fixed alongside it: its saved place could get overwritten by a short retry and force a full re-read of your history each run, which is what made a large import crawl and keep hitting the limit
• Your whole Claude history now imports, not just your most recent chats. LLMnesia reads your conversation list a page at a time by asking Claude for the chats older than the last one it saw. Claude's site was ignoring that and handing back the same newest page every time, and the fallback meant to cope with exactly this only ever re-read that first page before stopping. So on a busy Claude account, the import quietly kept only your most recent hundred or so chats and left everything older out of search, with nothing to say it hadn't finished. LLMnesia now recognises when Claude keeps returning the same page, switches to reading your list by position instead, and carries on from where it had reached, so even a large Claude history is imported all the way to the end.
• An interrupted import on Claude, Gemini, Perplexity or Grok now picks itself back up. Importing a long history pauses to wait out the AI's rate limits, and it can be cut short if you close the tab partway. On ChatGPT, LLMnesia already carried on from where it left off on your next visit
• The hints on the on-page corner indicator now fit their text. Small notes like "Right-click to continue in another AI" were squeezed onto two short lines inside a bubble that stayed stubbornly wide, leaving an obvious empty gap beside the words. Each hint now sits on one line in a bubble sized to match it, with no leftover space.
• Pages you left open now notice when LLMnesia updates. After an update, a tab you weren't actively using would keep showing a normal-looking indicator that had quietly stopped working, and only owned up once you clicked it or carried on chatting. Any open tab now spots the update on its own within a few seconds and switches to the "Reload to update" prompt straight away.
• Search-by-meaning setup now finishes on its own, and keeps up with new imports. Getting search-by-meaning ready processes your chats a few at a time in the background. It could get stuck partway, so the "getting ready" count would freeze (say at 718 of 2,143) and never move: a single chat that couldn't be processed halted the whole run, and even when it was going, it only made progress while something else happened to keep LLMnesia awake, so on a large library it stalled and waited for you to reopen search or press a button. And once it had finished, a fresh import of more chats was never picked up, so those new chats stayed unsearchable by meaning. All three are fixed. A chat that can't be processed is retried a few times and then set aside so the rest carry on
• The "chats saved" milestone note in the popup no longer trails off mid-sentence. Each time your saved-chat count passes a round number, LLMnesia shows a brief note in its popup. If you'd already added your email, or dismissed an earlier one of these notes, it still ended on a dangling "no card:" with nothing after it. It now reads cleanly: if you've already signed up it simply confirms your spot, and each new milestone shows in full rather than being silenced by an earlier one you dismissed.
• The "Try a handoff" setup tip now says where to click. It told you to open a saved chat and choose "Continue in…", without saying that option lives behind a right-click on the search pill. It now spells out the actual step.
• The search pill's progress bar no longer flashes right up against the "Saving…" text. When a save first appeared, the bar briefly rendered touching the label before settling into its normal spacing a moment later. It now sits in the right place from the start.

• It's now obvious when filters are hiding results — when a time range, platform, media, starred, or match filter is narrowing what you see, the search box shows a clear "Filtered" badge next to the result count instead of a small, easy-to-miss "· filtered" note. Clicking the badge opens the filters panel so you can see exactly what's active and clear it.
• A redesigned viewer for your saved chats — the page that opens a stored conversation (handy when the original has been deleted or is offline) has been rebuilt to read like a real chat rather than a data dump. It now shows which AI the chat came from with that platform's colour, lays the messages out as a clean back-and-forth with proper headings, lists, and code blocks, and carries a quiet reassurance that this is your own copy, kept on your device and readable even if the original is gone. Your search words are highlighted, there's a one-click "Copy transcript", and the technical details are tucked behind a "Details" toggle instead of crowding the top. When a chat isn't saved, a friendly note explains why and how to capture it, rather than a bare error.
• Landing on a deleted chat now explains itself — if you open a search result and the original has been removed from the AI's site, you're taken straight to your saved copy instead of a broken page. That copy now says up front why you're there ("We couldn't find this on Claude — it may have been deleted or moved"), so the switch reads as the product doing its job rather than a glitch.
• Jump straight to your saved copy from search — hovering a result now reveals a small "Open copy" button alongside the star and delete actions, opening your instant local copy without waiting on the live page. Alt+Enter or Alt/Option-clicking a result does the same.
• Gemini chats now save cleanly and in order — after a Gemini redesign, saved Gemini conversations came out jumbled: all your questions grouped first and then all of Gemini's replies (instead of the real back-and-forth), each reply stored twice, and stray "You said" / "Gemini said" labels mixed into the text. LLMnesia now reads Gemini's new layout turn by turn, so saved Gemini chats keep the right person on each message, in the right order, once each, without the leftover labels.
• Gemini no longer reports a problem while a page is still loading — occasionally LLMnesia checked a Gemini conversation a split second too early, before the page had finished drawing, and left the site marked as "couldn't read this chat" even though it read fine moments later. It now re-checks once the page settles, so a normal load no longer shows a false error.
• Pasted text in Claude chats is now searchable — when you paste a long block into Claude (a job posting, a CV, a chunk of code), Claude stores it as an attachment rather than as your typed message, and LLMnesia was skipping it — so a chat could be missing from search even when your search word was right there in the pasted text. That pasted and file-attached text is now saved and indexed along with the rest of the conversation. Your older Claude chats catch up on their own, too: while you're on Claude, LLMnesia quietly fills in the pasted text it missed before, a few chats at a time in the background — no re-import needed.
• "By meaning" results no longer jump above chats that actually contain your words — with search-by-meaning on, a closely-related chat that didn't contain any of your search words could occasionally rank above one that did. Chats that literally contain a search word now always come first, with meaning-based matches kept below them as related results.
• Long chats no longer go missing for a word buried deep inside them — to stay fast, LLMnesia keeps a capped list of words for each chat, and on a very long older conversation a distinctive word could fall off that list and become unsearchable. When a search comes up without a chat that matches all your words, LLMnesia now double-checks the full text of just those few long older chats and surfaces the one that genuinely contains them — without slowing down ordinary searches.
• Kimi chats now save with the right speakers — Kimi conversations were being stored with every message labelled "unknown" and your own questions dropped entirely, leaving a one-sided transcript. LLMnesia now reads Kimi's turns correctly, so saved Kimi chats keep both sides of the conversation with the right person on each message.
• Qwen chats now save with the right speakers and in order — Qwen conversations could land with turns mislabelled or out of sequence. LLMnesia now reads Qwen's turn markup directly, so saved Qwen chats keep the right person on each message in the order they were sent.
• DeepSeek saved chats now keep the answer, not the model's "thinking" — for DeepSeek replies that show a "Thought for…" reasoning step, LLMnesia was saving the private reasoning text and dropping the actual answer. It now captures the answer itself.
• Tables now display as tables in saved chats — a table in an AI's reply used to collapse into a run-on line of mashed-together cells. Tables are now kept intact and shown as proper rows and columns in the viewer (and when carried over via "Continue in another AI").
• Copilot bullet-point labels no longer go missing — in Microsoft Copilot replies, the bold label that starts each bullet (e.g. "Overall lifespan: 2–4 years…") was being dropped when the chat was saved, leaving stray bullets that began with a colon and read as one jumbled block. Those labels are now kept, so saved Copilot chats show properly formed lists — and the labels are searchable too.
• Search previews now show why a chat matched, not just your most common word. The one-line preview under each result picks the most relevant snippet from the chat, but on a word you use in nearly every conversation (say a project name that's everywhere in your history), it would land on a passage crammed with that word and miss the part that actually made the chat distinctive. So a chat you were sure you'd had could show up with a preview that gave you no way to recognise it. Previews now favour the rarer, more telling words in your search, so the snippet lands on the part that sets that chat apart.

• Add or edit a memory without leaving your chat — the context pill's picker now has a "+ Add a memory" option, and each saved memory has a small edit (✎) button, so you can jot a new memory or tweak an existing one right there in the chat instead of opening Settings. A memory you add from the pill is ticked and ready to drop into your message straight away.
• A setup checklist in the popup — a short three-step checklist (import your history, turn on memory, and try a "Continue in…" handoff) now guides you through the setup that makes LLMnesia genuinely useful, so the features that make search and memory worthwhile aren't left undiscovered. Each step links straight to where you do it and ticks itself off once done
• A gentle nudge to leave a review — once you've been using LLMnesia for a while and opening the chats you search for, a one-line prompt occasionally asks whether you'd leave a quick Chrome Web Store review to help others find it. "Leave a review" opens the store page in a new tab
• Choose how many results you see — the result count at the bottom of the search palette (e.g. "50 recent") is now clickable: tap it to pick 30, 50, 100, 250, or 500. It sets how many recent chats appear when the box is empty *and* how many matches a search returns — handy right after importing hundreds of chats, when the old cap of 100 made older ones impossible to reach. The same choice is available in Settings, and the two stay in sync.
• A one-click star filter in search — the search palette now has a ★ button right in its toolbar, so you can show just your starred chats in a single click instead of opening the Filters panel to find "Starred only". It lights up gold while active, and stays in sync with the filter in the panel.
• Break your usage down by year — the "Usage by platform" chart in Settings → Indexing status now offers per-year views (e.g. 2026, 2025) alongside the existing 7/30/90-day and all-time ranges, so you can see how your AI use shifted year over year.
• A "Your year in AI" snapshot — Settings → Indexing status now opens with a quick recap of your total activity this year: how many chats and messages you've had, how many different AIs you used, and a month-by-month sparkline with your busiest month highlighted. It appears once you've had at least one chat this year.
• A more helpful "no results" message — when a search genuinely finds nothing, the search box no longer just says results will appear here. It now tells you what actually happened right in the box — e.g. "Searched all 1,240 indexed chats — nothing matched", or "Still importing your chats" if an import is running — and suggests trying fewer or simpler words, so an empty search reads as "looked everywhere" rather than "maybe it's broken".
• The context pill is now one compact shape everywhere — instead of a wide bar on new chats (which could stretch across your message box) plus a separate card in existing conversations, it's always the same small two-row card: centered just above the message box on a new chat, tucked into the corner during an ongoing conversation. You can still drag it out of the way, and it can no longer sprawl across what you're typing.
• Clearer label on the popup's "Add a memory" box — the checkbox that used to read "Star (pre-select on new chats)" now reads "Use by default on new chats", which is what it actually does.
• The Settings page is reordered so the things you actually touch come first — the day-to-day sections (Indexing status, Memory, Starred, Snippets) now sit up top, followed by preferences (Search results, Storage limits), then "Report a Problem" — which used to be buried near the bottom — moved up next to Backup. The rarely-needed and irreversible bits (technical error log, then the "Danger Zone" delete-everything section) now sit at the very bottom where they can't be hit by accident.
• LLMnesia's occasional prompts no longer pile up — the handful of optional asks (stay in touch by email, support the project, leave a review) are now coordinated behind the scenes so only one can appear at a time and they're spaced several days apart, instead of potentially landing together over a busy week. Each still only shows when it's genuinely relevant
• You're no longer asked for your email twice — the "stay in touch" email prompt on the welcome screen and the one in the search box were separate, so closing the welcome tab without answering meant getting the same question again on your first search. The welcome screen now settles it, so it's asked once — with a later, optional reminder only after you've searched a good number of chats.
• The "index older chats" panel waits a beat — on a supported AI site it no longer springs up the instant the page loads
• One import nudge in the popup, not two — while the setup checklist is still walking you through your first import, the separate "Index older chats" strip stays hidden, so the popup isn't pushing you toward the same action in two places at once.
• A clearer, punchier welcome screen — the headline now reads "Search every AI conversation you've ever had", and the setup steps no longer contradict each other: step 1 is a straight one-click import of your history (no browsing needed), and step 2 explains that new chats index themselves automatically as you use AI. The stray "Open search" button — easy to click before anything was indexed, only to get no results — is gone
• The welcome Memory demo now matches the real thing — the little animation showing "Add my context" now puts the pill below the message box, exactly where it appears in a real chat, and the shortcut note no longer tells you to set up a shortcut that's already there — it just shows the key (Alt+Shift+M) to press.
• The welcome Memory demo is tidier — the "Add my context" pill no longer overlaps the message box, and the animation panel is shorter so it doesn't leave a band of empty space below.
• "Sync older chats" in the popup now helps you get to a supported tab — when you're not on a supported AI site, the import panel used to dead-end on "Open a supported AI tab to start." It now names which sites are supported and gives you a dropdown to pick one and open it in a new tab.
• The welcome checklist now ticks itself off as you go — the "Get your first search in 3 steps" list no longer just sits there as instructions. As soon as anything is indexed, steps 1 and 2 check off and you see a live count ("3 conversations indexed so far") climbing as your chats come in
• Clearer wording for the Memory step on the welcome screen — the button now reads "Add your first memory" instead of "Set up my memory", since Memory is already on by default — there's nothing to switch on, just your first memory to add.
• The popup "Get set up" checklist now keeps you in the popup — its steps no longer send you in circles. "Import your history" used to open a fresh AI tab even though the "Index older chats" panel right below does the same job in place
• A tidier "Index older chats" nudge — the reminder to import your history is now a single short line instead of two, so the popup reads faster.
• "Index older chats" now sits at the top of the popup until you've done it — bringing in your existing history is the most useful first move, so the import panel rises above the indexer stats until you've run it at least once, then drops back down so your live stats lead.
• A shorter, tidier popup — the three indexer controls (Indexing, Search pill, Memory pill) used to each take a full row with its own "Toggle" button. They're now compact chips on a single row — tap a chip to switch it on or off — so the popup is noticeably less tall.
• Removed a stray placeholder in Settings — the indexing status area no longer shows a confusing "No action yet." line before you've done anything
• Importing a backup now shows live progress — restoring a large backup used to sit on "Importing…" for a minute or more with no sign of life, so it looked frozen. It now counts up as it goes ("Importing … — 640 conversations so far…"), so you can see it working.
• The "Reload to update" note no longer spills out of its bubble — after LLMnesia updates itself, the little tooltip on the index pill asking you to reload the page now wraps neatly inside its box instead of running off the edge.
• Memories you add show up in Settings immediately — adding a memory from the popup or the chat pill now appears in an open Settings → Memory list right away, instead of only after you reloaded the page.
• The context pill no longer drifts off-screen when you drag it to an edge — dropping the pill near the side of the screen and reopening it could push part of it past the edge
• "Import from …" on the welcome screen now shows the import prompt right away — and it stays put — clicking, say, "Import from Claude" opened the site but the "Index older chats" panel arrived several seconds late and then vanished the moment the site finished loading (the site's own page navigation was tearing the panel down before you could click it). When you've explicitly asked to import, the panel now appears almost immediately and stays until you choose "Start now" or "Not now". (Passive visits to a supported site are unchanged — the panel still holds off a few seconds so it isn't the first thing that moves.)
• No more false "couldn't index this site" warning while a chat is still loading — opening the popup on a ChatGPT tab that hadn't finished loading could flash a red "this site couldn't be indexed" message, which then cleared itself once the page settled. The extension now waits for a page to finish rendering before deciding it has no conversation, so the warning only appears when a site genuinely can't be read.
• The popup's setup checklist now updates the moment you finish a step — saving your first memory or finishing an import used to leave the checklist looking un-done until you closed and reopened the popup (or visited Settings and came back). Both steps now tick off immediately.
• The popup's "Import your history" step now ticks off however you import — it used to complete only after finishing an "Index older chats" backfill, so importing from a site you use (e.g. Claude) or restoring a backup could leave it stuck as incomplete even though your chats were clearly indexed. It now counts any searchable history — matching the welcome screen — so you never have to import from an AI you don't use just to advance setup. (The separate "Index older chats" nudge still encourages importing your older chats when only a handful are indexed.)
• Fixed a fake-looking button in "Index older chats" — the panel used to show its title twice (once as a heading, once as a pill next to the Start button) and that pill was styled just like a button but did nothing when tapped. It's gone when idle
• The search pill's status text now lines up with its dot, and only shows a progress bar while something's actually happening — the label used to sit slightly off-centre next to the dot, and the pill kept an empty sliver of space reserved for the progress bar even at rest. The pill now stays snug when idle, and only grows to show a slim progress bar during an active save.

• A setup checklist in the popup — a short three-step checklist (import your history, turn on memory, try a "Continue in…" handoff) guides you through the setup that makes LLMnesia useful, with each step linking straight to where you do it and ticking off once done.
• Choose how many results you see — the result count at the bottom of the search palette is now clickable: pick 30, 50, 100, 250, or 500 to set both how many recent chats show when the box is empty and how many matches a search returns. Handy after importing hundreds of chats. Also in Settings, and the two stay in sync.
• A one-click star filter in search — a ★ button in the search palette's toolbar shows just your starred chats in a single click, without opening the Filters panel. It stays in sync with the panel's "Starred only" switch.
• Break your usage down by year — the "Usage by platform" chart in Settings → Indexing status now offers per-year views (2026, 2025, …) alongside the 7/30/90-day and all-time ranges.
• A "Your year in AI" snapshot — Settings → Indexing status now opens with a recap of your year: how many chats and messages you've had, how many different AIs you used, and a month-by-month sparkline with your busiest month highlighted.
• Fixes — the "Import from…" panel on the welcome screen now appears immediately and stays until you choose, and the popup no longer flashes a false "couldn't index this site" warning while a page is still loading.

• Index older Grok chats — open grok.com and click "Index older chats" to pull your past Grok conversations into search, the same way ChatGPT, Claude, Gemini and Perplexity already work. It runs quietly in the background
• See which assistants you use most — the Settings page now shows a "Usage by platform" breakdown under Indexing status, with a sorted bar for each assistant (ChatGPT, Claude, Gemini, and the rest) using that assistant's colour from search results. You can filter by how recently you used each one (last 7, 30, 90 days, or all time) and switch between counting whole chats or total messages, so you can answer questions like "which did I lean on most this week?" In the all-time view each row also shows when you last used that assistant, making it easy to spot ones you've drifted away from. Your most-used tools rise to the top and rarely-used ones fade at the bottom.
• Less clutter under the Settings indexing card — the technical "Last adapter errors" list has moved down into the "Technical error log" troubleshooting section where it belongs, and the internal "Migration status" line is now hidden unless a migration is actually in progress (instead of permanently reading "done").
• Clearer, livelier "Usage by platform" filters — the selected time range and metric are now highlighted in a bold blue so it's obvious which view you're looking at. Switching views smoothly re-ranks the platforms (most-used gliding to the top) while the bars grow or shrink and the counts tick up or down — no rows popping in or out. Platforms you didn't use in the chosen window dim down and settle at the bottom. Respects your system's reduced-motion setting.
• Tidier Settings while you're not importing — the "Index older chats" box on the Settings page now only appears while an import is actually running, where it shows live progress. Starting and pausing imports already lives in the extension popup (on a supported AI tab), so the duplicate, mostly-idle controls in Settings are gone.
• Search opens instantly from the shortcut anywhere — pressing the search shortcut on a non-AI page (any regular website) now pops the search box right there on the page, instead of waiting a couple of seconds and opening a separate search tab. The standalone search tab is still used on pages where Chrome doesn't allow it (its own settings pages, the Web Store, PDFs).
• "Usage by platform" only lists assistants you actually use — the breakdown under Indexing status no longer shows an internal "Anthropic Console" entry that isn't a supported assistant, matching the rest of Settings.

• Microsoft Copilot support — your Copilot conversations at copilot.microsoft.com now get indexed automatically as you browse, so they show up in search alongside your other chats. You can also continue a Copilot conversation in another assistant.
• Import Gemini and Perplexity history from the welcome page — the welcome screen now has "Import from Gemini" and "Import from Perplexity" buttons alongside ChatGPT and Claude. Each button opens that platform in a new tab, where you click "Index older chats" to pull your older history into search.
• Easier-to-tell-apart result colours — each AI platform now has its own distinct colour tint in search results, so you can spot at a glance which assistant a conversation came from. Several platforms that previously looked nearly identical (for example ChatGPT and Qwen) are now clearly different.
• More reliable first-run welcome — fixed a startup timing issue on brand-new installs that could make the very first session behave as if setup hadn't finished. The welcome page still opens exactly once after installing, and never again after updates.
• Readable Perplexity imports — imported Perplexity threads now show your actual question and the answer instead of raw behind-the-scenes data, so titles and previews read cleanly and search matches the real conversation.
• Correct dates on imported Perplexity chats — imported Perplexity threads now keep their original dates instead of all appearing as if created and updated at import time, so they sort correctly alongside your other conversations.

• Import older Gemini chats — Gemini now supports importing your existing chat history, so older conversations become searchable without opening them one by one.
• Helpful search reminders after import — after importing older chats, LLMnesia now gently reminds you to try search so you can check that your new index is useful. The reminder stops once you search.
• Clearer sign-in guidance — if you try to import ChatGPT history while signed out, LLMnesia now tells you to sign in and try again instead of looking like nothing happened.
• More private analytics — optional email signups no longer connect your email to product analytics, and analytics events are filtered to keep conversation text, titles, URLs, and other personal details out.
• A more consistent interface — the popup and search palette now share the same calmer visual style, so moving between them feels like one product.
• No duplicate Gemini results after import — imported Gemini chats and chats you later open normally now stay merged into one searchable conversation.
• Gemini chat dates stay accurate — imported Gemini conversations now keep their original dates instead of appearing as if they were created today.
• Search reminders work after partial imports — if an import stops early but still brings in chats, LLMnesia now treats it as successful enough to prompt you to try search. Empty draft chats no longer make the import look like it failed.

• Welcome page on install — installing the extension now opens a welcome page that helps you import your existing ChatGPT and Claude history so search works right away, instead of waiting for new chats to be indexed
• Guided first search — the welcome page now opens with a 3-step checklist: step 1 brings in your history (buttons that open ChatGPT/Claude where you click "Index older chats" to import, or automatic indexing on other supported sites), step 2 explains local on-device indexing, and step 3 shows how to search with a "Try a search" button. Import is now the primary action up front so your first search isn't run against an empty index, with a clear reminder that indexing stays on your device and no conversation data is uploaded
• Catch up across devices — after your first backfill, LLMnesia notices when recent ChatGPT or Claude chats started on another device, browser, or profile aren't searchable here yet, and offers to pull them in
• Welcome page: can now import from both ChatGPT and Claude — starting a ChatGPT import no longer greys out the Claude button (and vice versa), and the page no longer jumps to a "you can close this tab" state after the first import
• Continue in: formatting is now preserved — continuing a conversation into another AI keeps paragraph breaks, bullet/numbered lists and code blocks instead of pasting one unbroken wall of text
• Cleaner captured text — no "Show more/Show less" noise — ChatGPT's collapsible long-message toggle no longer leaks "Show moreShow less" into indexed messages or into conversations you continue elsewhere
• Continue in: more conversation context reaches the target — per-turn text limits have been tripled (15 k for regular turns, 8 k for code-heavy turns) and the most recent turn now gets up to 50 k characters, so the last assistant response is no longer truncated mid-answer
• Popup: clearer message on platform home screens — opening the popup on any supported site's home or new-chat page (Gemini, ChatGPT, Claude, Perplexity, etc.) now shows "Open a conversation to start indexing" instead of misleading "Indexed OK" or "Indexing failed" labels

• Index older chats — one-click indexing of your existing ChatGPT and Claude history into the search index. Runs invisibly in the background
• What's new strip in popup — appears once after each extension update, links to llmnesia.com/changelog
• Backup & Restore moved to full settings — export and import are now in the Settings page instead of the popup, giving them more room and removing the popup scrollbar
• Backfill: progress pill now clears on completion — the pill's progress bar and count (e.g. "156/157") no longer stay stuck after a backfill finishes
• Backfill nudge no longer fires while a run is in progress — the "index older chats" prompt now re-checks whether backfill started during the API sampling window before showing, preventing it from appearing mid-run
• Backfill: Cancel button — popup and options now show a Cancel button alongside Pause/Resume when a backfill run is active, stopping the run entirely rather than just pausing it
• Backfill progress in the pill — while a backfill run is active, the pill shows live progress ("Backfilling 42/150") with a fill bar
• Backfill: stale chats get refreshed — conversations already in the index are now re-fetched during backfill if the platform API reports a newer modification date, so chats you've continued using since LLMnesia first indexed them get their updated content and correct dates
• Backfill: imported chats now show their original dates — ChatGPT and Claude conversations indexed via backfill now carry the original creation and last-update dates from the platform API rather than today's date, so the search index reflects when chats actually happened
• Backfill: conversations now actually get indexed on ChatGPT — three root causes fixed: (1) history discovery skipped the first batch of conversations (ChatGPT server-side renders the first page and returns empty for API offset=0)
• Backup export saves as a single file — export now produces one .jsonl.gz file with a single save dialog instead of multiple small chunks
• Backup export works reliably on large indexes — fixed a silent failure where exporting after a backfill run would error mid-way
• Popup layout stable on open — conversation and message counts now render immediately as placeholders, preventing the Indexer Status panel from jumping once async data arrives
• Popup search hint streamlined — replaced the three-line shortcut reminder box with a compact single-line hint, reducing popup height

• Welcome screen — opens automatically on first install
• Indexing activation badge — once 10 conversations are indexed the extension badge shows "NEW"
• Analytics: zero_results_returned ghost-fires — search_performed and zero_results_returned now fire after a 400ms quiet period rather than on every settled 120ms search, eliminating same-millisecond timestamp collisions and intermediate zero-result events fired while the user is mid-keystroke
• Analytics: result_count property — search_performed now includes a result_count integer property, making zero-result queries queryable directly without correlating two separate events
• Analytics: identity not attached on second install — identified distinct ID is now mirrored to chrome.storage.sync on email submit and read back on init, so a second browser install (e.g. Chrome after Firefox) picks up the user's email automatically without requiring a separate email submission
• Continue-chat cross-tab injection — each "Open in" action now generates a unique request ID embedded in the target URL
• Continue-chat empty conversation guard — "Open in" now detects when there are no indexable user/assistant messages and shows a hint instead of opening a blank target tab
• Continue-chat message quality — conversation text sent to the target platform is now filtered: thinking-duration strings removed, data dumps skipped, code-heavy turns trimmed, max 16 turns kept
• AI Studio "Continue in" disabled — right-click "Open in" context menu is suppressed on AI Studio
• AI Studio manual-save UX — pill click now always saves
• Continue-chat missing assistant messages — "Continue in" was only carrying user messages when a platform's assistant responses had an unresolved unknown role (common for DeepSeek). A new middle pass now detects this case and re-runs with unknown roles treated as assistant, restoring full conversation export for all platforms.

• ↑/↓ arrow keys cycle through previous search queries from the search palette
• History dropdown: click the clock icon to browse and re-run previous searches, with per-item removal
• Guided onboarding flow for new installs
• "Continue In" tooltip now has a clearer label
• Fixed: search terms now save to history when the overlay closes before the debounce fires
• Fixed: snippet save popup no longer reappears immediately after clicking ✕
• Fixed: switching tabs in the palette now correctly filters/re-runs search
• Fixed: search palette no longer stays open when clicking outside in Chrome
• Fixed: Kimi, Qwen, and Mistral adapter reliability improvements
• Fixed: email capture prompt no longer reappears on every overlay open after install
• Beta: snippet saving — highlight text in any AI chat to save it for later

• Search results display improvements
• Popup layout adjustments

• Local-first full-text search across AI chat conversations
• Support for ChatGPT, Claude, Gemini, Deepseek, Perplexity, Grok, Mistral, Kimi, Qwen, AI Studio
• Keyboard shortcut to open search palette (Cmd/Ctrl+Shift+9, customisable)
• "Continue In" — open any indexed chat in a different AI platform
• Snippet highlight — jump from search result directly to the source message