Note ·

Strategic forgetting

External memoryKnowledge managementContext engineeringAgents

Breunig’s prediction for next year’s problem comes from running a personal agent for six months and watching its memory. The skills it writes for itself accumulate odd fixes; it solves one task one way and later rediscovers a different way for a similar task; the instructions and memories build up and begin to conflict. Very few people have run one of these harnesses long enough to see it, and he thinks it is the context problem of tomorrow: as the field leans on memory, it has to learn strategic forgetting, and nobody has.

The Hacker News thread had the sharpest line on the mechanism: old facts in a memory do not get deleted, they get contradicted, and the model picks whichever version is nearer in the prompt. One commenter answered with the Zettelkasten, append-only notes with periodic gardening to consolidate them, and said agents have a lot still to borrow from it; another runs a small ticket system where every finished task writes back what the next one should know. A third pointed to the view that imperfect human memory is adaptive rather than a defect, and that a machine with perfect recall of its own history may be worse at thinking for it.

This is the problem the enterprise context warehouse is designed around without saying so: read-mostly, provenance on every record, and the agent’s own learnings kept apart from the company’s record, in a pod of their own, where forgetting can be a policy rather than an accident. For my own notes the lesson is older. This site’s notes are my gardening; the private notes app is where the contradictions accumulate.

References

    Open the full search →