Check that a memory update changes the next session
The checked documentation separates profile-local files, staged writes and the frozen session-start snapshot. Verify a correction with a saved record and a fresh-session action. Model behavior remains untested.
- Published
- Oct 1, 2026
- Reviewed
- Oct 2, 2026
- Content updated
- Oct 30, 2026
- Applicable version
- Hermes memory documentation snapshot, 2026-10-02; runtime untested
- Platforms
- Supported Hermes host
Before you start
- For execution: a working model and one disposable profile used for every test session
Use a disposable Hermes profile. This is a reviewed test protocol, not a successful memory experiment: the first round checked the documentation and its source hash without running a model.
Write, correct, reopen
- Ask Hermes to save a fictional preference: for sample reports, use blue headings. Check the memory tool result or the actual saved record.
- In a new session, request a sample report and check whether the preference influences its output.
- Correct the preference to green headings and require the obsolete blue preference to be replaced or marked superseded.
- Use /new and repeat the task. Verify both the stored correction and the next action. Remove only the disposable test record afterward.
Check the effective store and session
Read MEMORY.md and USER.md under the same profile used by the agent. A write awaiting approval has not reached that store. Memory is injected as a frozen snapshot at session start: restarting a messaging gateway can resume the old session, so use /new after the correction. Do not approve unrelated pending writes.
Record each verdict
Keep the profile identifier, write result, superseded entry, fresh-session boundary and observed report choice. Mark each as pass, fail or not run. In this round the two live document checks passed with unchanged hashes; persistence, correction propagation and Agent action are all not run.
Common failures
If the preference appears only before /new, inspect whether it was saved at all. If old and new preferences coexist, resolve the conflicting record and test a fresh session again. Do not reset a real user memory store to make the test pass.
Evidence and limits
Official documentation checked
A linked source changed and this record awaits editorial review. Check the original before relying on these steps. (3)
- Documentation snapshot only; no Hermes executable version or model behavior was tested.
- A saved file, staged write and fresh-session action are separate acceptance checks.
Evidence files
Sources
Cite or read with tools
https://botclaw.tech/items/wf-memory-check