Recall: When AI Chatbot Memory Needs to Know What’s Still Valid

Understanding when information in a group chat remains valid poses a challenge for standard AI chatbots. While these bots can recall past messages, they struggle to differentiate between what was once stated and what is currently accurate. To address this, a new open-source Telegram bot named Recall was developed.

Consider a scenario where a team is discussing a deadline. During a conversation, one member says they will send a proposal by Friday. Two days later, another team member changes the timeline, stating they will send the proposal on Saturday. A standard chatbot would retrieve both messages, but without context, it cannot ascertain which statement is the active agreement.

Recall functions as a Telegram bot (@Recall_bot) and organizes group commitments into four distinct categories: DECISION, COMMITMENT, AMENDMENT, and COMPLETION. This structure is built upon three core components. Firstly, Walrus Memory on Mainnet serves as the immutable single source of truth, where every event is recorded as an on-chain blob.

Secondly, a local SQLite database acts as a rebuildable cache, with the ability to reconstruct the active state entirely from on-chain blobs if the cache is deleted. Thirdly, the Pure-Code State Resolver employs a deterministic algorithm to walk the amendment chain, superseding outdated records and avoiding the need for an LLM to determine the current state. Instead, the LLM receives active, non-superseded facts as input.

To ensure reliable extraction of structured JSON data without the constraints of paid APIs or quota limitations, Recall cascades through an automated chain of seven free models via OpenRouter. The key takeaway is that Recall can answer the question, "What was stored?" through its Walrus Memory component, while the State Resolver component answers, "What is still valid?"

In essence, Recall remembers the group's decisions and ensures that the bot remains informed of what is currently valid. By leveraging Walrus Memory, a rebuildable cache, and a pure-code state resolver, Recall effectively addresses the issue of maintaining accurate information within group chats.