Standard AI chatbots can recall past conversations, but retrieving stored data isn't the same as knowing which piece of information is currently valid.
We built Recall, an open-source Telegram bot, to solve this exact problem for group chats.
The Problem: Groups Change Their Minds
Consider a team discussing a deadline in a busy chat: — "I will send the proposal by Friday." (two days later) — "Actually, I'll send it on Saturday."
A simple vector or memory lookup retrieves both messages. But raw retrieval only surfaces historical facts—it cannot tell you that Saturday is the active agreement, leading standard models to hallucinate or get confused.
How Recall Works
Operating as a Telegram bot (@Recall_bot), Recall structures group commitments into four distinct event types: DECISION, COMMITMENT, AMENDMENT, and COMPLETION.
Its architecture rests on three core components:
- Walrus Memory on Mainnet (Single Source of Truth): Every event is committed to Walrus as an immutable blob. The local SQLite database is strictly a rebuildable cache: deleting it and running npm run restore-test reconstructs the active state entirely from on-chain blobs.
- Pure-Code State Resolver: We do not rely on an LLM to figure out current state. A deterministic algorithm in code walks the amendment chain to supersede outdated records. The LLM only receives active, non-superseded facts.
- OpenRouter Free Fallback Chain: To extract structured JSON reliably without quota bottlenecks or paid APIs, the bot cascades through an automated chain of 7 free models.
Key Takeaway
• Recall answers: "What was stored?" • The State Resolver answers: "What is still valid?"
"The group decides. Recall remembers. Walrus proves."
Top comments (0)