dsh-agent-memory

Memory plugins in the DSH ecosystem tend to swing between two extremes: one builds out JSON records, confidence scores, and state machines into a full machinery, while the other offers no cross-session consolidation at all. dsh-agent-memory takes a deliberately simple stance—memory is compression and retrieval, so the most honest form is a handful of living Markdown documents plus an agent that reads and writes them.

The plugin layers on top of official compaction, consuming already-compressed session summaries and using a cheap model to extract them into per-day files (episodic memory). A dream step, gated by 24 hours or 5 sessions, consolidates recurring facts into themed files like user.md and project.md (semantic memory). No raw logs are re-read; extraction runs in a single KB-scale call. Without an API key, reading, writing, and search still work—only the model-driven steps are skipped. The always-injected context is a deterministic memory catalog, costing zero extra model tokens; content is disclosed on demand via memory_search and memory_read, mirroring the DSH skill mechanism.

Six tools—search, read, catalog, save, correct, dream—cover the full lifecycle from retrieval to correction to consolidation. Users can open and edit files directly; agent writes are suggestions, and the human always has the final say. Ideal for developers who want an agent that genuinely remembers them across sessions while keeping every memory transparent and auditable.