Squirrow

A local-first memory manager for your AI chats. Save facts, decisions, preferences, snippets and file notes per project, then copy the ones you pick into ChatGPT, Claude or any other assistant. No account, no backend. Your data stays in your browser's IndexedDB.

Live at https://squirrow.thymeed.workers.dev. It is installable and works offline.

Install it as an app

• Chrome or Edge (desktop): open the site and click the install icon at the right of the address bar.
• Android (Chrome): open the menu and choose Install app.
• iPhone or iPad (Safari): tap Share, then Add to Home Screen.

After the first visit, the app opens without a connection.

Back up your data

Your entries live only in this browser on this device. Clearing site data, using a private window, or switching browsers or devices means you won't see them. Use Export JSON now and then to keep a backup, and Import JSON to restore it or move it to another device. Saved picks are included in both. Exports from before saved picks existed still import, with no picks. A banner asks you to back up when your last JSON export is more than 14 days old, or when you have never exported and your entries were first seen in this browser more than 14 days ago. Dismiss hides it for 7 days. Exporting JSON from the banner or the sidebar counts as a backup; Markdown export does not. The reminder and the date your entries were first seen are kept in local storage, and are never part of the export file.

What it does

• Projects with five kinds of entry: fact, decision, preference, snippet and file note. Each kind has its own color and icon, and the entry dialog explains the chosen kind with an example.
• Search matches the start of any word, so setup finds "Database setup" but base does not. Every search word must match, and title matches rank above body matches. Ties go to the most recently edited entry.
• Kind filter: click the kind chips under the search box to show only those kinds. Several chips can be on at once, and they work together with search.
• Copy for your AI chat: tick entries, see a token count, and copy a paste-ready Markdown block. Snippets are wrapped in code fences.
• Auto-pick: choose a model (or type your own context window), a share of it and a task, and Auto-pick fills that budget with the best entries, counted the same way as the copy panel. Entries matching your task words come first, then decisions, preferences and facts. Entries you tick by hand are kept, entries you untick stay out, and everything left out is listed so nothing disappears silently.
• Saved picks: name the entries you have ticked and save them for the current project. Load a pick later to tick exactly its entries again. Entries deleted since then are skipped, and Squirrow says how many were missing. Saving under a name that already exists updates that pick. Names are matched ignoring capitalization, so "core" and "Core" are the same pick.
• Import from chat: paste a chat summary and it is split into draft entries by its headings, bullets and code blocks, and each loose paragraph becomes a draft too. The lines of a hard-wrapped paragraph or bullet are joined, the first sentence becomes the title, and the details hold the whole text (a single short sentence is just a title). The kind of each draft is guessed from simple keywords. You can edit a draft's title, details and kind, or drop it, before saving the rest into the current project. A whole summary pasted inside one markdown, text or untagged code block is read as normal text; a block with another language stays a snippet, and a wrapper with same-length code fences inside is not unwrapped (use a longer outer fence). Nothing leaves your browser, and no AI is used.
• Keyboard first: / to search, n for a new entry, Esc to close.
• Safe updates: a new version waits until you click Reload, so it never swaps out the app mid-edit. When you do reload, every open Squirrow tab reloads, so finish typing in other tabs first.
• Export and import: JSON round-trips, and Markdown export is available. Import asks before it replaces your data.
• Backup reminder: a dismissible banner offers a JSON export when the last one is more than 14 days old, or when you have never exported and your entries were first seen more than 14 days ago. Dismissing hides it for 7 days.
• Beginner friendly: an empty-state guide with three steps, and a clear "Title is required" message instead of a silent failure.
• Pixel-art scenery (light and dark) with twinkling stars, shooting stars and hopping robots. The motion switches off when your system asks for reduced motion.

Develop

npm install
npm run dev      # local server
npm test         # Vitest (core logic, UI flows, service worker)
npm run build    # typecheck + production build into dist/
npm run format   # Prettier; CI runs format:check

The scenery images are generated: edit scripts/make-scenery.py and run python scripts/make-scenery.py from the repo root.

How it is built

React 19, TypeScript and Vite. Logic lives in src/core with no React imports: storage behind a Repository interface, ranking in search.ts, context building and token counting in inject.ts, and import parsing in portability.ts. Token counting loads the gpt-tokenizer o200k tokenizer in its own chunk just after the first paint, and falls back to a 4-characters-per-token estimate if that fails. Untrusted input is parsed in one place, parseSnapshot. See ARCHITECTURE.md for the design decisions.

Scalability

Measured timings for search, Auto-pick, export and the list from 100 to 50,000 entries, with the method and the limits of the measurements, are in docs/SCALABILITY.md. In short: up to about 1,000 entries per project the list, search and export stay under about 0.35 s, but Auto-pick with the real tokenizer already takes about 7 s at that size, and at 5,000 entries and above expect visible slowness.

Limits

• Token counts use the o200k tokenizer (gpt-tokenizer). They are exact for recent GPT models and approximate for Claude and Gemini, which use their own tokenizers. If the tokenizer fails to load, the app falls back to about 4 characters per token and labels the count "estimate". The tokenizer is about 1 MB gzipped and is cached for offline use on the first visit.
• The backup reminder is remembered per browser. Another browser or device will show its own reminder, and counts its own 14 days from when it first saw entries. Anyone who already has entries starts counting from their next visit.
• A saved pick stores which entries it holds, not their text. Editing an entry shows the new text in the pick, and deleting one drops it from the pick when the pick loads.
• Squirrow runs one version at a time per browser. If an older tab is still open, a newer tab shows a message and waits until that tab is closed or reloaded.
• Kind guessing for chat imports is keyword-based, so it will sometimes guess wrong. Well-known dotted names such as "Node.js" are read as tech names, not files. That is why you review the drafts before saving them.
• A chat import reads at most 200 drafts and 200,000 characters of pasted text, up to the last whole line before the cap, and the review screen says when a cap cut the paste. A single line longer than 200,000 characters is refused with a message, since no whole line fits.
• The list is not virtualized, which is fine for hundreds of entries. It would need changing for thousands.
• Tests cover src/core, the main UI flows and the service worker. Browser-specific behavior such as installing and clipboard permissions is checked by hand.

Privacy

Nothing is sent anywhere. There is no backend, analytics or account.

License

MIT