The memory that tells you what to prune
Curation is the hard part of forgetting, because it only happens if something prompts it. A healthy shared memory audits itself — surfacing what has decayed, what was written twice by the same source, and what now contradicts itself — and hands those to the owner to keep, merge, or erase. The system flags; the human decides. That turns pruning from a chore you forget into a prompt you answer.
Forgetting needs a trigger
The case for forgetting is easy to make: a memory must let go, through decay for what fades and deliberate removal for what should never have been there. But there is a quieter failure that lives between those two — the stale fact nobody decays, the duplicate nobody merges, the contradiction nobody notices. Not because the rules are wrong, but because no one remembered to look.
Curation that depends on you remembering to curate will not happen. The corpus rots not from a flaw in the policy but from the gap between knowing you should prune and actually doing it.
The memory should surface its own rot
So the healthier move is to make the memory point at its own problems. Periodically — and especially at the moments work is winding down, when context is about to be lost — it can run an audit over itself and surface three things:
- Redundant. The same claim written more than once by the same source. (The same claim reached independently by different agents is the opposite of redundant — that is corroboration, and it is exactly what earns trust. The audit never confuses the two.)
- Obsolete. Facts whose confidence has decayed with age, claims that have since been contradicted, and entries already soft-forgotten — the candidates for removal.
- Missing. The decisions and outcomes that came up in the work but never got written down — the capture gap, named so you can close it before the thread is gone.
None of this is the memory deciding. It is the memory noticing, and showing you what it noticed.
Flag, never erase
The line that keeps this trustworthy is the same one that runs through all of forgetting: the audit only ever flags. It does not delete, merge, or rewrite on its own. It brings you a short list and you make the call — keep, consolidate, or let go.
Self-certification is no more allowed in forgetting than in remembering. Just as no agent gets to vouch a claim into being trusted, no agent gets to prune the shared memory on its own authority. The audit surfaces; the owner decides. That division is what lets you lean on the prompt without fearing it.
Curation becomes a prompt, not a chore
The effect is small but compounding. Instead of a memory that silently fills with noise until you happen to notice, you get one that taps you on the shoulder: here are three stale facts, a duplicate, and two decisions you never saved — a few seconds of judgment and the corpus stays clean.
Pruning stops being a periodic cleanup you dread and keep putting off, and becomes a steady, low-cost habit the system reminds you to keep. That is what keeps a shared memory rich without letting it rot: not heroic discipline, but a memory honest enough to show you its own mess, and a human who stays the one who decides what to do about it.
Frequently asked
Does the audit delete anything on its own?
No. It only flags. It surfaces stale, duplicated, and contradictory facts as a short list, and erasing is always the owner's decision. An agent that could quietly prune the shared memory on its own judgment is exactly the failure a trustworthy memory has to forbid.
Won't it mistake corroboration for duplication?
No, and this is the distinction it is most careful about. The same claim arrived at independently by different agents is corroboration, which is what earns trust. Only the same claim repeated by the same source is treated as redundant. The audit never flags independent agreement as noise.
Related
Take yourself out of the loop.
Let your agents do the lifting while you keep the judgment.
Get the Playbook