Hermes Agent
GIGAZINE’s Hermes memory trial separates saving preferences from following them
GIGAZINE’s September 9 hands-on review found that Hermes could acknowledge a preference without saving it, then load a saved preference without consistently obeying it. Explicit memory-tool use and clearer formatting eventually worked; the memory graph still disagreed with the CLI.
A new GIGAZINE hands-on review tests a modest but revealing promise: tell Hermes how to structure a summary, start a new chat, and see whether the preference survives. The September 9 article reports several different outcomes that a single ‘memory works’ verdict would conceal. An acknowledgement was not initially a durable save. A later saved preference could be recalled in a new session without reliably controlling the answer’s format.
The reviewer eventually obtained the requested Summary heading, three bullet points and Details heading. That success followed explicit memory-tool instructions and a correction. The article also reports that the visual journey graph showed no memory marker even though the command-line listing found the stored entry. Those are distinct persistence, behavior and presentation observations, not one undifferentiated model failure.
The facts
- The inspected GIGAZINE page is dated September 9, 2026 at 22:35 and links the official persistent-memory documentation. - The first request asked Hermes to remember a three-line summary before a detailed explanation; the next session did not follow it. - Explicitly asking for the memory tool produced Saved to memory and Saved as user information messages. - A later new session could describe the saved format while still failing to use it in its answer. - The reviewer eventually obtained Summary, three bullet points and Details after making the requested format explicit. - The journey graph lacked a memory marker while hermes journey list in PowerShell showed the stored memory. - Official documentation distinguishes MEMORY.md from USER.md and describes session-start frozen snapshots.
Why it matters
Persistent state and reliable compliance are different product guarantees. A tool can correctly store and reload a preference while a model fails to apply it. Conversely, a convincing promise in the current chat can hide the absence of a durable write. The review supplies a useful acceptance-test shape for anyone evaluating a personal agent: inspect the saved record, open a genuinely new session, and check behavior on a task that makes the preference observable. A visualization should be checked against storage rather than treated as its authority.
Current
Inspected on 2026-09-10. The Hermes Agent stable-release baseline is v2026.9.7 published 2026-09-07T22:17:01Z. The main source was GIGAZINE hands-on review dated September 9 when captured. Release metadata is a version boundary; an open proposal, a merged mainline change and a packaged release remain different states.
Evidence
The primary evidence is GIGAZINE — Hermes memory hands-on, September 9 (https://gigazine.net/gsc_news/en/20260909-hermes-agent-memory/). Supporting context comes from Hermes — persistent-memory documentation (https://hermes-agent.nousresearch.com/docs/user-guide/features/memory/); Hermes Agent — current official README (https://raw.githubusercontent.com/NousResearch/hermes-agent/main/README.md). These pages were inspected directly. Related project records explain the source cluster; they are not independent replications of contributor test claims.
Operator take
Use a harmless formatting preference for a repeatable memory test. Verify the record independently before judging a model’s response, and separate a request to inspect memory from permission to change it. Keep the test’s model and runtime version with the result. The article’s attribution of quirks to a free model is the reviewer’s interpretation, not an isolated causal experiment.
Caveat
This is one reviewer’s environment, with machine-translated English text and no independent reproduction by illmethinks. It does not establish failure rates across Hermes models or versions. Current official memory documentation is baseline context, not proof that the reviewer exercised every current policy.
This is one reviewer’s environment, with machine-translated English text and no independent reproduction by illmethinks. It does not establish failure rates across Hermes models or versions. Current official memory documentation is baseline context, not proof that the reviewer exercised every current policy.