Most "AI memory" problems aren't storage problems. They're filing problems.

Most "AI memory" problems aren't storage problems. They're filing problems.

I've been building a multi-agent setup at home where several AI harnesses (Claude Code, Codex, cron jobs, autonomous agents) all work on the same system. The obvious question came up fast: where does their memory live?

The tempting answer is "one database, everything goes in." That's the trap.

What actually worked was splitting memory into three stores, each authoritative for a different kind of fact:

Git holds durable, curated truth — the plans, the decisions, the standards. Anything that has to survive, be diffable, and be reviewed before it counts.

An episodic memory API holds session history — what each agent did, decided, and learned this time round, so the next one inherits context instead of starting cold.

Local harness memory is just a private scratchpad. Convenient, but never the source of truth, because no other agent can see it.

The whole design exists to prevent one failure: putting a fact in the wrong store. A decision that belongs in git rotting in a local cache. Session noise polluting the durable record. Each store has one job, and "promote it to the right place" is a first-class rule.

Funnily enough, this is just good filing discipline — the same instinct that keeps a homelab sane — applied to machines that remember.

Curious how others are handling memory across multiple agents. One big store, or a split like this?

#AI #LLM #MultiAgent #Homelab