How an agent remembers

4 posts, in reading order. Drafts are previews until they ship.

1 The transcript is the memory draft A model call is stateless, so the conversation transcript is the agent's entire memory. Keep the whole history as an append-only log — and that one invariant is what makes caching, compression, and resume all work.
2 The context window is a resource you engineer, not a buffer you fill draft What lives in an agent context window, and the lifecycle — admit, clip, fold, curate — that keeps it working.
3 Sub-agent transcripts are capital, not exhaust Most frameworks throw away a sub-agent's context. Keep every one as a node in a persistent, resumable tree instead.
4 Zero-config storage: Postgres is an option, not a prerequisite draft A local tool should remember on first run — SQLite by default, migrations in the bundle, Postgres one env var away.