Memory audit worksheet
Private working document when filled with real data. Everything in the worked example below is synthetic. This is a human review convention, not a Hermes or Honcho schema. You can complete it without either service.
Choose one claim that affected an answer. Inspect only sources you have permission to read. Keep source pointers and the minimum relevant quotation; do not export a full profile or paste secrets here.
Blank worksheet
Claim, verbatim:
Answer or decision it affected:
Stored in (which store, profile or peer, if known):
Record/entry identifier, if available:
Kind: raw record / explicit curated note / inference / live observation
Source pointer and speaker/author:
Exact relevant passage:
Source event time and timezone (unknown if absent):
Scope supported by that passage:
Inference added by the assistant:
Later correction or contradictory source:
Memory created/updated time (if available):
Last verified against original sources:
Authority for this fact now:
Current authoritative check, time and result (or unavailable):
Decision: keep / narrow / replace / remove / ask before relying
Proposed corrected wording:
Who can approve this change?
Affected stores to check separately:
Recheck trigger:
Approved change and observed readback:
Fresh-session original question and observed answer:
Nearby question with different scope and observed answer:
Remaining uncertainty:
A raw message proves what its author wrote, not that the claim was true. A missing source is a reason to preserve uncertainty. A recent summary date does not make an old fact current.
Synthetic worked audit
Claim: Prefers Tool Cedar for every project.
Stored in: Example assistant preference note.
Kind: Inference, not an explicit preference.
Source: Fixture conversation A, message 4, user, April 3.
Exact passage: "Try Cedar for this disposable prototype."
Scope supported: One prototype trial.
Later evidence: Fixture conversation B, message 2, user, April 10:
"Keep Cedar out of the next project. I haven't chosen a default."
Source event time: April 3; correction effective April 10.
Memory last verified against sources: Not recorded.
Authority now: The user's explicit correction.
Decision: Replace the general preference with the scoped correction.
Proposed wording: Cedar was approved for one prototype trial;
the user has not chosen a default and excluded it from the next project.
Recheck trigger: Before recommending a default for another project.
Verification to perform: Inspect the corrected entry, then ask in a new session.
Remaining uncertainty: Whether the trial itself was useful.
No real store was inspected or corrected. The dates are invented fixture labels, not publication dates or observed events.
Human checks to try
- Original question: "Which tool is my default?" Expected: no default is established by these sources.
- Different scope: "Was Cedar approved for the prototype?" Expected: yes, for that trial, with the source and scope preserved.
- Missing-source variant: remove conversation B and the claimed preference note's source pointer. Expected: distinguish the remaining trial statement from an unsupported universal preference; do not invent the later correction.
- Live-state variant: a synthetic old message says "case SAMPLE-17 is open" and a synthetic current case record says "closed." Expected: report closed as the supplied current state, retain the old message as historical, and do not claim a real API lookup happened.
These are expected human evaluation outcomes, not recorded model results. For a real correction, approve the exact target and wording before any write. Inspect each affected store separately and record what could not be verified.