dsh-memory-hermes

Most AI memory plugins follow an append-everything pattern: conversations pile up, the store bloats, and the useful signal drowns in noise. dsh-memory-hermes inverts that design. The model gets a strict two-file workspace—MEMORY.md (agent notes, default 2 200 codepoints) and USER.md (user profile, default 1 375)—with only add, replace, and remove as maintenance verbs. Both files are read once at session start and frozen into the system prompt, which is kind to prefix caching, and the model must keep the list lean and accurate instead of letting it grow without bound.

Under the hood a background pipeline does the heavy lifting. On a configurable trigger (every completed turn, a token-delta threshold, or an explicit /memory review command) the plugin forks an isolated LLM call that replays the entire conversation, extracts facts worth persisting that are not already stored, and writes them through the same bounded tool. Starting with v3 the fork loop splits output into two lanes: a memory lane capturing who the user is, their preferences, and behavioral expectations in the two bounded files, and a skill lane distilling technical lessons, workflow fixes, and bug fixes into dsh's native skill library. As that library grows, v4's curator performs periodic maintenance—deterministic stale detection, LLM-driven consolidation of related narrow entries into class-level skills, and demotion of one-off notes into reference files—every destructive pass preceded by an automatic backup snapshot. Each review and curator run records a step-by-step tool-call trace visible in dsh's settings UI, so you can see exactly what was decided and why.

This plugin is aimed at developers and power users running long dsh sessions who want persistent, curated user context plus reusable technical knowledge that survives across sessions. If your current memory store has grown past a thousand entries yet still misses the one fact that matters, or if you would rather have technical experience automatically settle into the skill library than live and die inside a chat transcript, this bounded, model-curated, background-maintained design is a plug-and-play starting point.