Why AI Agents Need Smarter Forgetting, Not Bigger Memory Windows

A software engineer argues that the real challenge in AI agent memory is not retention but selective forgetting, after observing multiple production agents degrade over time. The agents failed not because they lacked storage, but because they accumulated contradictory and outdated facts that buried relevant information. Most current implementations function as append-only logs with vector search, which cannot distinguish between a current fact and one that has since been superseded. The author proposes treating agent memory like a cache rather than a diary, using mechanisms such as explicit fact supersession and salience-based decay to keep memory stores accurate. Without these data-modeling changes, retrieval quality will continue to worsen as agents run longer, a flaw that cannot be fixed through prompt tuning alone.