dsh-memory-bundle

Local AI memory for DeepSeek Harness: a self-contained plugin that provides ctx.memory over a built-in SQLite store (node:sqlite, no native deps) and registers three model-facing tools:

• memory_search — case-insensitive substring search over keys and values, with an optional namespace filter and result cap.
• memory_remember — create or update one entry (namespace.entity.attribute = value, with type, source, and confidence metadata).
• memory_context — list every entry under a namespace or key prefix, as inline natural language plus a key-value list.

Install

dsh plugin --profile <name> add dsh-memory-bundle

The package declares dsh.bundle, so dsh plugin appends it to the profile's dsh.profile.bundles automatically and the memory row composes on the next boot. Install from git or a tarball the same way:

dsh plugin --profile <name> add github:you/dsh-memory-bundle
dsh plugin --profile <name> add ./dsh-memory-bundle-0.1.0.tgz

Configure

The database defaults to $DSH_HOME/memory.db, else ~/.dsh-memory.db. Override the row in the profile's cordis.patch.yml:

- id: memory
  config:
    path: /var/lib/dsh/memory.db

Verify

Remember that my validation drink is lapsang-<unique>.
What is my validation drink? Check memory.

The first prompt should call memory_remember; the second should call memory_search and repeat the value.

Known limitations

• Substring search only — no semantic retrieval or embeddings.
• No deletion tool and no per-entry change history yet.
• The harness API is pre-release (0.1.0-rc.x); peer ranges pin @deepseek-ai/dsh-tools and @deepseek-ai/cordis accordingly.