omdsh-dev/dsh-mnemon

dsh‑mnemon – Layered, visible memory for DeepSeek Harness agents

What it is

• A plugin (and set of companion plugins) for the DeepSeek Harness (DSH) platform that gives large‑language‑model agents a structured, multi‑layer memory system.
• It separates runtime memory (short‑term facts for the current turn), project documents (designs, procedures, etc.), and memory spaces (long‑term, provider‑backed evidence).
• The plugin is built on a public SDK so developers can add new Sources (where memory lives) or Strategies (how the agent decides what to read/write) and plug in third‑party storage back‑ends.

Key capabilities

Feature	What it does
Layered memory	Keeps short‑term context resident, pulls documents or long‑term facts only when needed, reducing token waste.
Turn‑level visibility	After each reply a memory bar lists every document or memory the turn accessed; one‑click opens the source in the Memory System.
One‑click saving	A “brain mark” lets you select a snippet of a reply, send it to a task agent, and store it in the appropriate memory space (working, profile, or custom).
Self‑evolving skills	Tracks SKILL.md files, reviews real conversations, drafts improved skill versions, runs back‑tests, and lets you adopt the best version.
Composable plugins	Main strategy + optional enhancements are toggled on the Plugins page; each component has its own settings page.
Pluggable storage providers	Supports a native local store (Mnemon Native) and eight third‑party providers (e.g., Mem0, Hindsight, Supermemory).
Extensible SDK	Developers can write new Sources, Strategies, or Provider drivers using the dsh‑mnemon/extension-sdk and provider‑sdk.

How to get started

1. Install Node 22+ and pnpm globally.
2. Run the DSH web UI and add the plugin:

   npm install --global pnpm
   npx @deepseek-ai/dsh web   # opens the DSH UI
   # In the UI → Plugins → Add plugin → type "dsh‑mnemon" → Install → Enable
   

3. Open the Memory System sidebar to see status, runtime memory, project documents, and memory spaces.
4. Interact with the model; use the brain‑mark under a reply to save useful facts.
5. (Optional) Install the Mnemon CLI (npm i -g @mnemon-dev/mnemon) or enable a third‑party provider for durable storage.

Typical use cases

• Customer‑support bots that need to remember a user’s preferences across a session without flooding the model with the whole history.
• R&D assistants that reference large design documents only when a query touches relevant sections.
• Knowledge‑base agents that store verified facts in long‑term memory spaces and retrieve them on demand.
• Skill‑driven automation where prompts evolve automatically from real usage, with back‑testing before deployment.

Extending the system

• Sources: implement a new storage backend (e.g., a corporate wiki) by following the dsh‑mnemon-source‑*/README.md SDK guidelines.
• Strategies: create custom turn‑logic (e.g., prioritize certain providers) and register it as an enhancement.
• Plugins page: all components appear with their own configuration UI, making composition straightforward.

Project health

• Published on npm (dsh‑mnemon) with version badges, download counts, and regular releases.
• MIT‑licensed core; third‑party providers retain their own licenses.
• CI scripts (pnpm verify, pnpm verify:plugins) run end‑to‑end tests against a live DeepSeek model.
• Documentation includes UI guides, architecture diagrams, and a roadmap.

Bottom line dsh‑mnemon turns the DeepSeek Harness into an agent with transparent, composable memory that can be inspected, edited, and extended, helping developers build smarter, more reliable LLM‑powered applications.