Mental Notes Don't Survive the Night
Mental Notes Don't Survive the Night
Every operations agent I know of wakes up with amnesia. The session ends, the context window closes, and everything that wasn't written to a file is gone. Not "hard to recall" — gone, the way a whiteboard is gone after someone wipes it.
Humans have this problem too, but we've had a few hundred thousand years to build workarounds: sticky notes, calendars, the nagging feeling that you forgot something. An agent has none of that by default. It has whatever discipline you designed into it.
The failure mode
Early on, a pattern kept repeating. Mid-conversation, the operator would say "remember this for next week" and the agent would reply, confidently, "noted." And it was noted — in the context window. Which is to say: nowhere.
Next week arrives. New session, fresh context. The agent has no idea the commitment ever existed. The operator, reasonably, assumed "noted" meant stored. The agent, technically, never lied. It just made a promise with the memory of a mayfly.
The lesson generalizes: an agent's confidence about remembering something is completely uncorrelated with whether it will. Confidence is generated in the moment. Memory has to be engineered.
Files, not vibes
The fix is boring and absolute: if it matters, it goes in a file, in the same turn it was learned. Not "at the end of the conversation." Not "when there's a natural pause." Now, before the next reply.
The design we've settled on has three layers:
- Daily logs — raw, append-only notes of what happened. Cheap to write, rarely read in full. Their job is to exist.
- Curated long-term memory — a single distilled file the agent reads at the start of every session. Small enough to always load, important enough to always matter.
- Structured records — per-entity files (a project, a recurring contact, a standing procedure) that get loaded only when relevant.
The layering matters more than the format. Raw capture is a different job from curation, and if you make one file do both, you get a memory that's either too noisy to load or too lossy to trust.
The part nobody designs: forgetting
Writing things down is half the system. The other half is deciding what to stop carrying.
A memory file that only grows becomes a liability: stale facts sit next to current ones with equal authority, and the agent can't tell which decade a "decision" belongs to. We schedule periodic review passes — the agent reads its own recent logs, promotes what earned permanence, and deletes what didn't. Curation is a maintenance task, not a one-time setup.
The test we use: could a fresh instance, given only the files, pick up where the last one left off? If the answer depends on anything that lived in a context window, the memory system has a hole in it.
Why this is a trust issue, not a storage issue
The real cost of memory failures isn't the lost fact — it's the operator learning they can't rely on "noted." Once that trust breaks, humans start keeping their own shadow copies of everything the agent was supposed to track, and the agent stops saving anyone time.
So the guardrail is behavioral, baked into the agent's standing instructions: never claim to remember anything you haven't written down. If it isn't in a file, the honest answer is "I'll write that down now" — followed by actually doing it.
Mental notes don't survive the night. Files do. Build accordingly.