dsh-workspace-memory

When you run multiple conversations in the same workspace, the hardest part is rarely model capability—it is context drift. The working habits, project decisions, and terminology constraints you settled in one session simply vanish by the next. dsh-workspace-memory splits shared context into two layers: reusable behavioral rules live in AGENTS.md, while stable facts, decisions, and terminology are kept in .dsh-memory.md. Both are ordinary Markdown files in the session working directory, readable, reviewable, and versionable without a database, embeddings, or any cloud service.

Its signature design is the approval gate. When the model identifies a piece of feedback worth keeping long-term, it does not write silently. It first reads the full target document, integrates the smallest coherent edit into the relevant section, and then presents a focused diff for line-by-line review. The file is written only after the user confirms, and the observed file version is re-checked at write time—stale proposals based on an older version are rejected outright. On DSH Web you get a dedicated review card with line numbers, colored additions and removals, and collapsed unchanged sections; other clients fall back to Markdown.

If you use DeepSeek Harness to run parallel conversations in one workspace, need to persist team working conventions and key project decisions, and want to avoid bolting on a memory service or vector database, this plugin is built for you. It keeps concurrent conversations aligned, makes every inferred write transparent and opt-in, and keeps the entire memory layer local, inspectable, and free of network calls.