dsh-superlocalmemory

A thin DeepSeek Harness (dsh) bundle that connects SuperLocalMemory to the harness over stdio MCP.

SuperLocalMemory is a local-first memory layer for AI agents: one memory that any harness speaking MCP can share. This bundle contains no memory logic. It is one configuration row for DeepSeek Harness's official MCP client plugin (@deepseek-ai/dsh-mcp-client) that launches slm mcp, so SuperLocalMemory's tools show up in the harness as mcp__slm__<tool>.

Prerequisites

1. DeepSeek Harness (npx @deepseek-ai/dsh web, see the harness README).

2. SuperLocalMemory installed so that slm is on your PATH:

   npm install -g superlocalmemory
   # or: pipx install superlocalmemory
   slm --version
   

Install

dsh plugin --profile web add dsh-superlocalmemory

Until the package is on npm, install from a clone instead:

git clone https://github.com/qualixar/dsh-superlocalmemory.git
dsh plugin --profile web add ./dsh-superlocalmemory

Restart dsh web. The bundle is appended to the profile's bundle list and its patch inserts one MCP client row:

- insert:
    - id: memory-superlocalmemory
      name: '@deepseek-ai/dsh-mcp-client'
      config:
        serverName: slm
        transport: stdio
        command: slm
        args: [mcp]

If you prefer not to install a bundle, copy that row into $DSH_HOME/profiles/<name>/cordis.patch.yml, or pass the file once with dsh web --patch ./cordis.patch.yml.

Verify

1. Open Settings > Plugins > Plugin list. A mcp-client entry with id memory-superlocalmemory should show as Mounted and Enabled.
2. Start a session. Initial discovery is asynchronous, so wait for the mcp__slm__... tools to appear.
3. In one session ask the agent to remember a unique fact, then in a new session ask it to recall that fact. The harness's own memory guide describes this write, fresh-session recall check.

You can also confirm the server on its own: slm mcp should start and wait for MCP messages on stdin.

Changing the command

If slm is not on the PATH that the harness sees, set command to the absolute path of the executable (which slm).

Licence

This bundle is MIT licensed (see LICENSE). It only contains configuration. SuperLocalMemory itself is a separate project under AGPL-3.0-or-later with its own licence terms; installing it is your choice and is governed by its repository.

SuperLocalMemory is built by Varun Pratap Bhardwaj as part of the Qualixar AI Reliability Engineering research initiative. This bundle is not affiliated with or endorsed by DeepSeek.