dsh-memory-hermes

Most AI assistant memory schemes follow a "pile it up" model: every exchange gets appended to a growing store, entries balloon, and signal-to-noise degrades over time. dsh-memory-hermes takes the opposite approach. It gives the model a bounded two-file workspace (MEMORY.md for agent notes, USER.md for user profile) with hard codepoint limits, and only three maintenance actions: add, replace, remove. At session start the files are read once and frozen into the system prompt—prompt-cache friendly—and the model is forced to keep a sharp, compact list of persistent facts rather than an ever-growing dump.

The real depth is in the background pipeline. On a configurable trigger (every turn, token-delta threshold, or manual command), the plugin forks an isolated LLM call that replays the full conversation, extracts facts worth keeping that are not yet stored, and writes them with the same add/replace/remove semantics. v3 splits the output into two lanes: a memory lane that records who the user is, their preferences, and behavioral expectations in the bounded files, and a skill lane that distills technical lessons, workflow fixes, and bug fixes into dsh's native skill library under $DSH_HOME/skills/. As the skill library grows, v4's curator runs periodic maintenance—deterministic stale detection, LLM-driven consolidation of prefix clusters into class-level skills, demotion of narrow entries to references—each destructive pass preceded by an automatic snapshot. Every review and curator run is logged with a step-by-step tool-call trace visible in dsh's settings UI.

Built for developers running long dsh sessions who need persistent user context and reusable technical knowledge across sessions. If you are tired of a memory library that grows to thousands of entries yet still misses the key fact, or if you want technical experience to automatically sediment into reusable skills instead of living and dying in chat transcripts, this bounded, model-curated, background-maintained approach is an install-and-go starting point.