A model call remembers nothing. Everything an agent seems to remember, somebody stored on purpose: a fact for a later conversation, or the state of a run that must survive a crash. Those are two different jobs with two different failure modes. This site starts from one request that forgets and climbs, one idea at a time, to memory kinds, stores, run identity, checkpoints, safe resume, write policy, retrieval, shared memory, security and tests, asking one question at each step: who wrote this down, and who is allowed to rely on it?
An underwriting run fails right after it fetched the roof assessment. Start over and the provider is called twice and the underwriter may see two recommendations. Resume carelessly and the run trusts a stale fact. Memory engineering keeps the facts an agent should remember apart from the saved state that lets a run continue, and makes both safe to rely on.
Working, episodic, semantic and procedural memory; history against summaries against facts; files, relational tables, documents and vector stores.
run_id, trace_id, attempt and task ids kept apart; state machines, workflow checkpoints, event sourcing and where replay stops.
Idempotency keys, the transactional outbox, compensation and deduplication, so a resumed run never repeats a write that already happened.
What may be written, with which provenance and expiry; ranking by relevance and recency; shared memory, tenants, personal data and deletion.