I kept watching my agents forget things that mattered. Not edge cases. The central limitation that stopped them getting more useful as their job grew.
I run three production agents: Maia, a marketing agent with 150+ tools; hive; and Deepak, a chief-of-staff agent. Each one failed differently. Maia would analyse a brand's voice in one session, then start from scratch the next because it came in through a different channel. Serving two brands, it let the casual one bleed into the formal one. It treated a six-week-old brief as current.
The off-the-shelf memory systems assume a single user, a single channel, a single purpose. All three of my agents break all three assumptions at once, so those systems produce interference, source amnesia, and temporal amnesia by design.
So I stopped treating memory as a feature and started treating it as a pathology. I named the ways an agent forgets: anterograde (it never persists), interference (one client's context leaks into another's), temporal (facts with no timestamps), source amnesia (a fact with no owner), and context-binding (the right memory exists, but the retrieval query never matches how it was stored). Each one has a direct, observed manifestation, and each one gets a measurable test.
Engram is the framework that comes out of that: memory as typed, importance-weighted, semantically searchable records, with extraction running as fire-and-forget async work after each turn so it never blocks the agent.
The thesis is simple. Memory is not something these agents happen to use. It is the mechanism that turns an impressive demo into something genuinely useful.