The four-stage pipeline
Conceptually, Rethread is a small, sandboxed pipeline that runs entirely inside your browser:
- Capture — observe what's said on the page
- Extract — distill conversations into structured memories
- Organize — buckets, tags, folders, search index
- Recall — pick the right memories and inject them into a new chat
Optional Sync sits behind the whole pipeline as a zero-knowledge encrypted layer. Let's walk through each stage.
Capture: reading what's already on the page
Rethread is a Manifest V3 Chrome extension built on WXT. On each of the six supported AI platforms (chatgpt.com, claude.ai, gemini.google.com, grok.com, www.perplexity.ai, chat.deepseek.com), a content script attaches a lightweight DOM observer.
When you send a message and the AI responds, the observer reads the new turns from the rendered conversation — exactly as they appear to you. No network requests are intercepted, no API tokens are read, and the page is not modified beyond a small Recall button and a context preview UI.
A debounced job picks up the new turns and forwards them to the extraction worker.
Extract: distilling raw chat into structured memories
Raw chat logs are noisy and rarely useful as long-term memory. Rethread runs each new conversation chunk through a lightweight extractor that produces structured "memories" of four types:
- Fact — durable information about you ("I'm a backend engineer using Go and PostgreSQL")
- Preference — stylistic / behavioral choices ("I prefer functional over OOP")
- Decision — outcomes you've committed to ("We chose Cloudflare Workers for the v2 backend")
- Context — current situational state ("My project uses Next.js 14 with App Router")
Each memory is timestamped, tagged with its source platform and conversation, and assigned a confidence score. You can edit any memory afterwards, change its type, add tags, or move it to a different bucket.
Organize: buckets, folders, tags, snapshots
A flat memory list works for the first 50 entries. After that you need structure. Rethread gives you several layers:
- Buckets — color-coded top-level groups (e.g. "Work", "Side project", "Personal")
- Folders — nested groupings inside buckets
- Tags — free-form labels for cross-cutting concerns ("typescript", "client-feedback")
- Snapshots — immutable copies of any conversation at a point in time, with word-level diff against later snapshots or the live conversation
- Time range filters — slice by "last 7 days", "this month", custom range
Bulk operations let you multi-select memories or whole conversations and re-bucket, re-tag, export, or delete them in one sweep.
Recall: Selective Recall and context injection
The most-used surface in Rethread is Selective Recall. Press Alt+Shift+R on any of the six AI platforms (or click the floating Recall button) and a panel opens with:
- A search box that ranks memories by relevance to your current page / current draft
- Filters by bucket, tag, type, time range, source platform
- Per-memory toggles for inclusion + per-memory edit-in-place
- A live token estimate that updates as you check / uncheck memories — so you never blow your context window
- A choice between Summary mode (compact bullet list) and Detailed mode (full memory bodies)
- Per-memory feedback (helpful / hide) to teach Rethread which memories are signal vs. noise
Click Inject and Rethread drops a clean context block into the AI's prompt area. From the AI's perspective, this is just regular text — no hidden hacks.
Optional: zero-knowledge cloud sync
Pro users can layer end-to-end encrypted sync on top of the local pipeline:
- You set a passphrase. Rethread runs it through PBKDF2 (200,000 iterations) and HKDF to derive an AES-256-GCM key, all in your browser.
- Memories, conversations, snapshots, buckets, and templates are encrypted on-device before upload to Cloudflare R2.
- Our backend (Cloudflare Workers + D1 + R2 + Durable Objects) stores opaque ciphertext + IVs + minimal metadata (record ID, type, timestamp, byte size).
- To sync to another device: install Rethread, activate the same license, enter the same passphrase. Decryption happens locally.
- Lose the passphrase and the cloud copy is permanently unreadable — by design. There is no recovery path.
Read the full deep dive on the encrypted sync architecture →
What Rethread does not do
- It does not call OpenAI / Anthropic / Google / xAI APIs on your behalf.
- It does not modify the AI's responses.
- It does not phone home with telemetry, analytics, or usage statistics.
- It does not request broad host permissions — only the six AI domains.
- It does not require an account, login, or OAuth flow for the free tier.
Performance considerations
Rethread is engineered to feel invisible. The DOM observer uses a debounced batched listener; extraction runs in a background script worker; the IndexedDB store is indexed on the fields you actually query (text, type, tags, buckets, timestamp); and the Recall UI uses virtualized lists for libraries with thousands of memories.
Even on a $200 Chromebook with a 5-year-old battery, Rethread adds a fraction of a second to page interactions you wouldn't notice in a blind test.
See it for yourself.
Free to install, no account required, works in 60 seconds.
Add to Chrome — Free