dsh-memory

There is a gap in the DeepSeek Harness ecosystem: an agent that wants to remember a fact or a note has no durable, cross-session write channel. session_search can only read back past conversations; todo_write tracks task state; skills are read-only static docs; settings belong to the operator. None of them let the model actively persist knowledge. dsh-memory fills exactly that gap with a lightweight, model-editable, key-based memory store — letting the agent write, read, and retrieve what it wants to keep, as naturally as calling any other tool.

Four tools form the core: memory_set idempotently upserts an entry under a stable key (with optional tags), memory_get reads one entry or lists all, memory_delete removes by key, and memory_search performs case-insensitive substring matching across keys, values, and tags with tag-based filtering. All records are routed through the built-in ctx.storageDomain seam into the deployment's configured backend (json or sqlite), so there is no manual file management and no need to stand up an external MCP memory server.

This plugin suits DeepSeek Harness developers who want their agent to carry memory across sessions. It ships as plain ESM JavaScript with no build step and no native addons, is MIT-licensed, and exposes only a handful of config knobs (defaults: 1000 entries, 120-char keys, 10000-char values) — install and go.