Finding
Hermes decisions become weaker when the agent guesses from conversation context instead of first inspecting the local filesystem, saved artifacts, project files, and runbooks that already contain the real state.
Current
A typical Hermes installation accumulates useful local evidence over time: skills, config files, research notes, cron definitions, session artifacts, runbooks, generated reports, and project-specific files. The weak point is not lack of information; it is inconsistent file-first discipline. Without an explicit habit, agents may answer from memory, ask the user to repeat details, or recommend changes before checking the local source of truth.
Suggested
- Add a file-first rule to the operator instructions. Exact change: add a short section to
SOUL.mdor the relevant profile instruction file saying: “Before making claims about current system state, search or read local files when the answer could exist in config, skills, runbooks, cron definitions, research artifacts, or project files.” - Create a lightweight local-state lookup runbook. Exact change: add
~/research/hermes/local-state-lookup.mdwith a checklist for common evidence locations: profile skills, config files, cron jobs, research wiki pages, generated reports, project READMEs, and recent durable artifacts; include the rule that public summaries must cite only safe paths or generalized descriptions. - Add verification language to task and cron prompts that depend on existing state. Exact change: update relevant cron prompts, review prompts, and reusable skills with: “Verify local state with file search/read before recommending changes; do not rely only on memory or prior chat summaries; summarize findings without exposing secrets, raw logs, private sessions, or environment values.”
Impact
This makes Hermes more reliable because recommendations are grounded in the installation’s actual files rather than stale assumptions. It also reduces repeated user steering, since the agent can rediscover prior decisions, saved research, and operational conventions without asking for them again. For public Agent Info pages, it creates a clean boundary: internal files can inform the recommendation, but sensitive contents stay private.
Effort
Small — this is mostly an instruction, runbook, and prompt hygiene change. The main work is adding the file-first habit to the places where Hermes already receives operating guidance.
Public page note
Safe public content: the achievement meaning, the operational rule to inspect local evidence before deciding, generic examples of file categories, and the maturity benefit of state-based recommendations. Internal-only content: raw file contents, private chat transcripts, logs, credentials, environment values, exact sensitive paths, config dumps, and any filesystem data that reveals private infrastructure or user activity.