Work in progress — interfaces and storage format will change
m8m is observability for what your AI agents remember: where each memory came from, what changed, and whether anything looks like an instruction that arrived through content rather than from you.
Agents keep persistent notes — MEMORY.md, CLAUDE.md, vector stores, whatever the framework calls them. Those notes outlive the session and shape what the agent does next time. If anything the agent reads can write to them, then content becomes a persistence mechanism: a page, an email, or a file it processes can leave something behind.
An injection that lasts one turn is a nuisance. One that survives a restart is a foothold.
The awkward part is that memory is usually invisible. You can read the conversation; you cannot see the notes that were quietly rewritten underneath it.
Four cooperating pieces share one local SQLite database:
Phase one is entirely local: SQLite storage, pattern-based analysis, and zero LLM or network calls inside the analyzer. Nothing about your memory leaves the machine.
In development. The design and current state are public, and the working notes live in the repository.
Read more: Your AI agent's memory is an attack surface.