AgentRecall
A correction-first persistent memory system for AI coding agents that stores structured records of user corrections and measures whether an agent actually stops repeating a mistake, rather than only measuring memory retrieval.
AgentRecall is a governed corrections ledger paired with a measurement instrument for AI agent memory. Every time a user corrects an agent, the correction is stored as a structured record with severity, evidence, and outcome tracking, persisting across sessions, projects, and agent restarts. Unlike most memory tools, which only test retrieval, AgentRecall tracks whether a correction actually changed the agent's later behavior: each correction accumulates a retrieved count, and every time the agent meets the same situation again, the outcome is logged as heeded or recurred.
It installs as an MCP server for Claude Code, with a corresponding SDK and CLI, and a per-session loop that loads context at session start, records corrections as they happen, and compounds what was learned at session end. The project publishes its own measured numbers, including a correction-capture recall rate and heed rate from a dual-blind audit on its own corpus, rather than only claiming the system prevents repeated mistakes.
It is written in JavaScript under the MIT license.