aqsk-BLG/dsh-memory

README excerpt

dsh-memory Layered file memory for DeepSeek Harness — one plugin row that brings global snapshot injection, workspace notes, session recall, post-compaction flush, and a bundled usage skill. The design was informed by a survey of memory systems across the agent ecosystem (OpenClaw, Hermes, Codex, WorkBuddy, and others) and follows the file-first doctrine they share: memory is plain files on disk, a small curated layer is always in context, and detail lives in a deep layer reached on demand. What you get Layer Files Behavior Global memory $DSH HOME/USER.md , $DSH HOME/MEMORY.md Injected into every session as a frozen snapshot (4,000 / 1,500 char budgets, truncated with a usage header) Workspace memory /.dsh/memory/MEMORY.md , /.dsh/memory/YYYY-MM-DD.md Never injected; read on demand with the file tools. Daily logs are append-only; logs older than 30 days are distilled into MEMORY.md Session recall all session transcripts The session search tool (self-contained queries; the calling session is always excluded) Compaction flush — After a successful compaction, a reminder asks the model to persist important context before it leaves the context The bundled memory runtime skill is the com…

Category