Hermes Agent
A Hermes user finds long sessions can hide the value of persistent recall
A Raspberry Pi user reports nine unsolicited session searches across 1,990 messages and only five Hermes sessions. The account, shared on Hacker News September 9, argues that daily restarts preserved context rather than creating the boundaries needed to exercise long-term recall.
A hands-on account by heyjonny gives Hermes' self-learning pitch a concrete usage question: what happens when the user almost never starts a new session? After moving from OpenClaw to Hermes on a Raspberry Pi 4 with Telegram, the author inspected a copied state database with Claude's help and reported nine voluntary session_search calls across 1,990 user messages. Those messages were concentrated in five sessions over 18 days, including one lasting more than a week.
The post's explanation is that shutting the Pi down each evening did not create a new conversation the next morning. Hermes resumed the old session, keeping much of the relevant material in immediate context. This is an account of one user's workflow, not a controlled test of memory quality. It nevertheless separates continuity from recall in a way that a feature checklist cannot. The HN submission dates its public circulation to September 9; the fetched article itself does not expose a publication date.
The facts
- The author says the installation ran on a Raspberry Pi 4 and used Telegram after a switch from OpenClaw. - The reported database review found nine self-initiated session_search calls among 1,990 user messages. - The account covers five sessions over 18 days, with one session extending beyond a week. - Powering down overnight was followed by resuming the same session, according to the post. - The author reports that MEMORY.md and USER.md entries were usually compressed into shorter abbreviations rather than removed. - Official Hermes documentation distinguishes bounded persistent memory files from the session-search path into earlier conversations.
Why it matters
A low recall-tool count does not automatically mean memory is broken. The agent may already have the information in its current context. Conversely, a long-running session can make a system seem to remember well while never testing whether useful knowledge survives a genuine boundary. This matters for both evaluation and cost: the author describes sessions becoming long and expensive, but supplies no controlled cost comparison or generally optimal reset interval.
Current
Inspected on 2026-09-11. The Hermes Agent release baseline is v2026.9.7 published 2026-09-07T22:17:01Z. The main source was undated author account circulated on HN September 9 when captured. Release metadata bounds availability; proposals, merged mainline changes and packaged releases are not interchangeable.
Evidence
Primary evidence: heyjonny — I think I used the Hermes Agent wrong (https://heyjonny.dev/posts/i-think-i-used-the-hermes-agent-wrong/). Supporting context: Hermes — persistent memory and session-search guide (https://hermes-agent.nousresearch.com/docs/user-guide/features/memory/); Hacker News — heyjonny memory account, September 9 (https://news.ycombinator.com/item?id=49628740). These pages were inspected directly. Linked documentation establishes the product contract; it does not independently reproduce contributor or author experiments.
Operator take
Test memory with a deliberate, reversible session boundary and a specific fact that should persist. Check whether the answer comes from a durable note, a searched conversation or immediate context. Avoid adopting a reset schedule solely to increase tool-call counts; successful recall is the goal, not busier telemetry. The distinction also keeps a user-experience report from turning into a claim that all Hermes users are operating the product incorrectly.
Caveat
The statistics are author-reported and the database was not independently examined. The post does not name an exact Hermes version or establish causal performance results. HN is a circulation timestamp, not proof of article authorship date.
The statistics are author-reported and the database was not independently examined. The post does not name an exact Hermes version or establish causal performance results. HN is a circulation timestamp, not proof of article authorship date.