Conversation history is not automatically long-term memory. Keep recent turns for immediate coherence, use reviewed summaries for compression, and write durable semantic memories only under explicit product policy. Every memory needs subject, scope, provenance, confidence, sensitivity, validity, and deletion behavior. Retrieval combines semantic similarity with user, session, time, and policy filters.
Do not use embeddings as the only identity key. Filter by authenticated subject and scope before ranking.
Some memories require user confirmation before approved. A newer correction should supersede rather than silently overwrite provenance.
Test two users with similar preferences and prove no cross-subject retrieval. Correct a preference, expire it, delete it, and resolve an old reference. Compare summaries with source turns for lost negation or dates. Inspect what happens when no memory is relevant.
Design a consent-aware memory policy for [assistant use case]. Classify memory types, eligible facts, prohibited data, confirmation, confidence, validity, corrections, retention, deletion, retrieval filters, and user controls. Include cross-user and stale-memory tests.
Verification contract: A user must be able to inspect, correct, and delete eligible durable memories, and deletion must propagate through every serving path.