dsh-memory-canvas

Running multi-step agent tasks means tool logs pile up fast, filling the context window while what you actually need to keep is just the skeleton of the task structure. dsh-memory-canvas tackles exactly that: it offloads bulky tool output to local refs files, leaves a single Mermaid canvas in context to track task nodes, and lets the agent pull the full body back by node ID when needed. No daemon, no vector database, no ports to babysit. Everything lives in a local .dsh-memory directory, making it simple, auditable, and dependency-free.

The plugin exposes three tools the agent calls explicitly: memory_offload writes a title and body into refs and returns a short ID while updating the canvas; memory_recall fetches stored content by ID (truncated at 8k with a note); memory_canvas prints the current Mermaid Markdown in full. It does not auto-intercept logs. The agent decides when to offload and when to recall, keeping full control over context pacing.

Ideal for developers running long-chain tasks inside DeepSeek Harness who want the lightest possible way to manage the context window. It is not a persona pipeline, not a vector search engine, and not a daemon-backed service. It is a minimal references-plus-canvas memory layer that slots into the tool layer of your agent.