Max Chronicle — memory for AI agents, so they stop asking me the same question twice
An open-source memory layer my agents use every day: an MCP server, a CLI and a SQLite event store with hybrid recall and its own eval harness.
700+
Tests in CIRAG
Hybrid retrieval: full-text, vector, timeMCP
Server for any agentMIT
Open sourceThe Challenge
Agent sessions forget. Every new session started from zero: re-reading files, re-asking questions, re-making decisions that were already made — and sometimes contradicting them. With several agents and models working on the same projects, context loss became the most expensive bug in the system.
The Approach
Built a local-first memory that any agent can use through MCP or a CLI:\n\nEvery event, decision and checkpoint goes into a SQLite event store. Each stored assertion carries its kind — observed, decided or assumed — so an agent knows what is a fact and what is a guess.\n\nRecall is hybrid: full-text, vector and time-based, and it keeps working offline on full text when the embedding service is down. A built-in eval harness measures recall quality, so changes are judged by numbers, not by feel.\n\nHandoffs are explicit: an agent records a checkpoint (what's done, what was verified, what's open, what's next), and the next agent — or the next model — picks it up without the previous conversation.
Orbital diagram showing a multi-agent system with 7 specialized AI agents. The central Dispatcher routes tasks to Job Scout, Content Writer, Code Reviewer, Researcher, Data Monitor, and Telegram Bot. Click any node to see details and connections.
The Result
It runs my agent workflows every day, including the sessions that built this site. 700+ tests run in CI on Python 3.11–3.13. It is open source under MIT — see the repository link above.
