Introduction¶

When used across sessions, DeepSeek Harness (DSH) agents often lack the ability to reuse historical experience, causing them to repeatedly solve the same problems. The dsh-self-memory plugin automatically records session traces and solutions, provides offline retrieval and closed-loop write-back capabilities, and helps agents directly apply past experience in subsequent conversations.

Prerequisites¶

Before using this plugin, ensure the following environment requirements are met:
- Node.js version >=22
- DSH version @deepseek-ai/dsh 0.1.x

Install and Enable¶

Install the plugin from a local directory. Hot mounting is supported.

dsh plugin --profile web add <本目录>

After installation, verify the plugin status with the following command:

selfmem_status

The response should include information such as plugin=v0.3.0. After code updates, you can use dev_reload_package dsh-self-memory or restart DSH to complete the setup.

Core Capabilities¶

The plugin mainly addresses the following five issues:

1. Multi-keyword relevance: Weighted keyword matching improves retrieval relevance.
2. Detail suppression to enhance generalization: When recording, it automatically generates abstractProblem; during retrieval, it matches both the raw text and the generalized text.
3. Fast retrieval over large case sets: It uses an inverted index to filter candidate sets and supports large-scale records.
4. Conflicting results under different preconditions: Record-level preconditions and solution-level appliesTo and conflictsWith support precondition-based filtering and conflict annotation.
5. Unified keyword normalization: It includes a built-in synonym table, and retrieval and storage use a unified normalization path.

Usage¶

The plugin completes closed-loop memory management through a series of tool commands.

Automatic injection and daily closed-loop¶

The plugin enables automatic capture and automatic injection by default. After each user message is sent, the system automatically retrieves historical solutions and injects the context via a <selfmem_memory> fence. When solving new problems, the agent should follow the workflow: “Check records first → If found, try it → Record the root cause of success/failure → If not found, solve independently → Finally write back”.

Command Tools¶

• selfmem_status: View plugin status and scale statistics.
• selfmem_search: Retrieve existing solutions, with context precondition filtering.
• selfmem_record: Record or append solutions; supports multiple solutions, preconditions, and keyword weights.
• selfmem_hit / selfmem_fail: Mark a solution as successful or failed and record the root cause.
• selfmem_import: Import solutions from external sources.
• selfmem_trace: View automatically captured session traces.

Usage Examples¶

Record a solution with preconditions:

selfmem_record {"problem":"断网时如何安装插件","solution":"解压本地包 → dsh plugin add <目录> → pnpm install → dev_install_package 热挂载","keywords":{"plugin":1,"network":0.6},"appliesTo":{"network":"no"}}

Sample precondition-aware retrieval:

selfmem_search {"query":"如何识图","context":{"model_vision":"no"}}

Configuration¶

You can adjust plugin behavior by overriding the profile’s cordis.patch.yml.

Configuration example:

- insert:
    - id: self-memory
      name: dsh-self-memory
      config:
        autoRecall: true
        recallBudget: 800
        captureEnabled: true

Notes¶

• Zero runtime dependencies: The plugin runs entirely locally, does not rely on external services, and supports offline use.
• Retrieval mechanism: Current retrieval is based on keywords, generalized text, and inverted indexes; vector semantic retrieval is not yet supported.
• Record distillation: Session traces are captured automatically, but they are not automatically distilled into records; manual tool invocations are required to write them back.
• Precondition definitions: For multiple solutions to the same problem, appliesTo still needs to be manually specified when recording to distinguish them.
• Storage limits: Storage is currently a single-file JSON store; consider migrating to SQLite after reaching tens of thousands of records.

Conclusion¶

dsh-self-memory provides a basic cross-session memory capability for DSH agents through structured records and closed-loop write-back. For more details, refer to the project documentation:

• Project directory
• GitHub repository