Making corrections stick

Someone should be able to correct a memory from the same place they see it. They need to point at the wrong value, give or pick the replacement, and see what source or authority that type of record needs.

Don't edit history quietly. Mark the old record superseded or invalid, create the corrected one, link the two and record who made the change. Current-state retrieval then leaves the old value out, while anyone reviewing history can still see why it showed up in earlier tasks.

Deletion is a separate thing from correction. Privacy or retention rules might mean removing the record, the source content and any indexes built from them. Keep track of derived artefacts like embeddings, caches and summaries so a deletion can actually reach them. A tombstone can keep a minimal, non-sensitive identifier so the same content doesn't get re-ingested by accident, as long as your retention policy allows that.

Changes to sources need to flow through. If a document is replaced, or its access is narrowed, find the memories that came from it. They might need to be revalidated, restricted or invalidated. A source link that no longer resolves should lower trust in the record and create a maintenance task, instead of sitting there as a dead reference nobody notices.

Users should be able to challenge an inferred preference without having to supply a replacement. Sometimes removing the inference is the right outcome. Then you need to stop the system recreating it straight away from the same old evidence, so keep a scoped suppression record or change the extraction policy.

Corrections also need to clear caches and active-session context where you can manage it. If the durable store is updated but an agent keeps working from a cached old value, people get a confusing delay where the fix seems not to have worked. Return a correction version that clients can check before doing anything consequential.