dsh-plugin-memory

The plugin-based architecture of DeepSeek Harness (DSH) allows developers to extend functionality through plugins. In agent development, knowledge is discarded after a conversation ends. The dsh-plugin-memory plugin uses the Session Distillation mechanism to transform completed conversation turns into persistent knowledge entries and register them as Skills for use in subsequent sessions.

The plugin is maintained by 16480000 and is licensed under MIT. It distills question-and-answer interactions and tool calls into structured data, stores them in a local directory, and exposes them to the model through memory-<slug> skills.

Installation¶

Install via npm (recommended, no build permissions required):

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

Install via GitHub (from source):

dsh plugin --profile <name> add github:16480000/dsh-plugin-memory

When installing from GitHub using pnpm ≥10, add the allowBuilds configuration to the profile’s pnpm-workspace.yaml:

allowBuilds:
  dsh-plugin-memory: true

Install via a local tarball:

pnpm pack
dsh plugin --profile <name> add ./dsh-plugin-memory-0.1.0.tgz

Configuration¶

The plugin supports overriding default settings through the profile’s cordis.patch.yml. The main configuration items are as follows:

Example override configuration:

- id: memory
  name: dsh-plugin-memory
  config:
    autoDistill: false

Core Features¶

1. Automatic Collection
   On every turn/end with reason completed, the plugin collects the question-and-answer content and tool call results from the conversation.

2. Distillation and Merging
   The plugin invokes an auxiliary model to distill the collected content into a strict JSON format, including title, answer, sources, and an optional merge_with field. If the model determines that the current entry has the same topic as an existing entry, returning merge_with triggers a merge call; if merging fails, the new entry is kept and overrides the existing entry.

3. Persistent Storage
   Entries are persisted in the <DSH_HOME>/memory directory. Each entry generates a Markdown file <slug>.md and an index file _index.json. Index writes are atomic (temporary file + rename).

4. Skill Registration
   The plugin registers a memory skill provider and offers memory-<slug> skill candidates (Rank 550, ranked below hand-written skills). Loading the skill returns the corresponding entry Markdown content.

5. Service API
   The plugin registers the ctx.memory service, which provides the following methods:

   • list(): Returns the current memory index.
   • readEntry(slug): Reads the Markdown content of a specific entry.
   • settled(session): Waits for all pending distillation tasks in the session to complete.

Notes¶

• Missing security policy: The plugin has no content security policy, and distilled answers are stored as-is.
• Filter limitations: The incomplete answer filter is marker-based and may produce false positives or false negatives.
• Scope isolation: All sessions share one memory directory; isolation by workspace or agent scope is not supported yet.
• Management mechanism: The directory lists all entries, with no session-level limits or size-based eviction mechanism.
• Dependency requirements: A DeepSeek Harness build that includes the purpose: 'memory' vocabulary and the memory/* event mechanism is required. Older versions can still load the plugin and distill content, but some functionality will be missing.

Conclusion¶

dsh-plugin-memory provides basic Session Distillation capability for DeepSeek Harness. It is suitable for scenarios that require building long-term conversation context or a knowledge base. Review the source code and license before installation.