Memory system

A memory system is any mechanism that lets an agent remember something after the session ends. Since the model is stateless and the context window is thrown away, remembering has to be built: the harness writes facts to files or a database during a session and loads them back into context at the start of the next one. That is all memory is, at every scale, from a MEMORY.md to a vector store.

Two failure shapes. One, no memory at all: every session re-learns your conventions, your quirks, and the thing you told it yesterday. Two, memory that grows without curation, so each session starts with a wall of stale notes competing for attention. A memory that says the tests take five minutes when they now take thirty is worse than no memory, because it is trusted.

Keep memory small, factual, and dated. Store decisions and preferences, not transcripts. Put project-wide facts in AGENTS.md where every session sees them, and keep personal or volatile notes in a memory file the harness loads on demand. Review it occasionally the way you would a runbook. Anything the repo already records, such as code structure or history, does not belong in memory; the agent can read the primary source.