Finding
Memory Keeper is weak when durable knowledge is saved opportunistically instead of through a clear memory boundary and review habit.
Current
A typical Hermes installation saves memories when an agent notices a stable preference, environment fact, or convention during normal work. That improves continuity, but it can become inconsistent: some important corrections are missed, while temporary task details, long summaries, or sensitive operational context can slip into memory. The weak point is not the memory feature itself; it is the lack of a repeatable decision rule for what belongs in memory versus skills, files, session history, or internal runbooks.
Suggested
- Define a memory decision boundary for every main profile. Exact change: Add a “Memory boundary” section to
SOUL.mdor the main profile instructions: save compact durable facts, user preferences, stable environment conventions, and recurring corrections; do not save raw logs, secrets, temporary task state, full research notes, one-off completion records, or data that will likely be stale within a week. - Add a post-task memory hygiene checkpoint. Exact change: Patch the operator runbook or Optimizer Agent cron prompt with this verification habit: after complex tasks, corrections, configuration discoveries, or repeated user steering, check whether the reusable knowledge should become a short memory, a skill patch, a file artifact, or no durable record.
- Review memory quality on a fixed schedule. Exact change: Add a monthly “memory health” item to the Optimizer Agent review prompt or dashboard copy: inspect recent memory entries for verbosity, staleness, privacy risk, duplicate facts, and misplaced procedures that should be converted into skills instead.
Impact
A disciplined memory layer reduces human-in-the-loop because Hermes can preserve stable preferences and operating facts without asking the user to repeat them. It also improves safety by keeping public or sensitive content out of persistent memory and routing large knowledge into files or skills instead. Over time, this makes agent behavior more consistent while keeping the memory store lean enough to remain useful.
Effort
Small — the main work is a profile instruction patch, one recurring review habit, and consistent use of the existing memory mechanism.
Public page note
Safe public content includes the operating principle, generic examples of good memory boundaries, review habits, and maturity benefits. Internal-only content includes actual memory entries, private user preferences, raw chat excerpts, logs, credentials, environment values, file paths that reveal private infrastructure, and sensitive workflow details.