AI memory is broken. We built one that forgets.

Every agent framework has the same problem with memory: it doesn't forget. Context windows reset between sessions. RAG and vector DBs store everything with equal weight and grow until they're noisy. So when your project changes direction two weeks in, the AI still pulls up week-one decisions like they're current. What this actually looks like Week 1: You tell the agent "we're using React for the frontend." Week 2: You switch. "Moving to Svelte, React bundle is too big." Week 4: You ask "what's our frontend stack?" A normal retrieval system hands back both answers. React and Svelte sit side by side with equal weight. Nothing in the system knows one replaced the other. So the agent might reference React, Svelte, or some confused mix of both. We kept running into this while building agent tooling, and it became clear the issue isn't retrieval quality — it's that these systems have no concept of time or obsolescence. The numbers We ran a 4-week simulated project through both systems. 24 ev