Introduction

DeepSeek Harness (DSH) uses a plugin-based architecture, allowing developers to extend agent capabilities. When building agents, the reliability of memory is a core challenge. Traditional memory approaches often rely on storing raw documents. This not only consumes a large amount of context, but also easily leads to retrieved information lacking contextual relevance or containing biases.

DeepBrain is different from document storage: it stores judgments—conclusions distilled from meetings, interviews, and lectures. Each judgment comes with an evidence chain, traceable back to the original recordings and cross-validated across different contexts. The dsh-deepbrain plugin brings DeepBrain’s “organizational second brain” capability into DSH, enabling agents to read these judgments, cite quotes verified verbatim, query personnel histories, and leverage DeepBrain’s analytical methods.

Plugin Positioning

dsh-deepbrain is a memory plugin for DeepSeek Harness, maintained by qiuyiwu1989-star and licensed under the MIT license. Its core value is to provide DSH agents with a structured memory layer: judgments, evidence chains, and verbatim-verified quotes.

The plugin does not store raw documents; it stores only processed judgments. All calls are live calls and are not cached.

Installation and Enablement

Before installing, ensure that DeepSeek Harness is installed.

  1. Install the plugin:
    Use the following command to install the plugin into the specified profile (default is web):
    npx -p @deepseek-ai/dsh dsh plugin --profile web add github:qiuyiwu1989-star/dsh-deepbrain

Restart the profile after installation completes.
  1. Configure environment variables:
    DeepBrain requires an account and an API Key. It is recommended to set them through environment variables first, instead of writing them into configuration files (prevents secrets from leaking into logs or screenshots):
    export DEEPBRAIN_API_KEY=lj_live_xxx

Core Features

The plugin registers 6 MCP tools with DSH. The agent does not need to invoke the 27 raw tools provided by the server. The plugin is responsible for shaping them into LLM-friendly forms, preventing the model from making blind selections among too many tools and causing context explosion.

  1. deepbrain_brief

    • Purpose: When you are unsure what to ask, describe the current situation first and let DeepBrain decide what key information you need to know.

  2. deepbrain_topics

    • Purpose: Browse the index or retrieve the summary for a topic. This is the end point for most writing and analysis work.

  3. deepbrain_quotes

    • Purpose: Retrieve attributable quotes, each verified verbatim against recordings.

  4. deepbrain_person

    • Purpose: Query the personnel list or a person’s historical positions. Recommended for use before negotiations or meetings.

  5. deepbrain_ask

    • Purpose: Ask a specific question and obtain an answer with sources. Slow queries automatically wait for the result.

  6. deepbrain_call

    • Purpose: Escape hatch. If the five tools above cannot cover the need, directly call any DeepBrain MCP tool.
    • Note: If the tool parameter is empty, it lists all tools that the current API Key is allowed to call.

Configuration Example

Add the following content to the profile configuration in DSH:

- insert:
    - id: deepbrain
      name: 'dsh-deepbrain'
      config:
        endpoint: 'https://shennao.zaowuyun.com/api/mcp'
        apiKey: ''              # 留空表示读取 DEEPBRAIN_API_KEY 环境变量
        timeoutMs: 120000       # 多步骤检索可能耗时较长，建议设置超时
        passthrough: true       # 设为 false 则不注册 deepbrain_call

API Permissions and Call Limits

DeepBrain API Keys are scoped. The server prunes the available tool set according to the permissions of the Key.

  • brain.ask: Read judgments, quotes, personnel information, and the method library. This is the basic permission required for most use cases.
  • transcript.read / transcript.write: Read or write raw recording transcripts.
  • brain.propose: Write judgments (these judgments enter the inbox and require manual review).

Write operations are performed only when explicitly requested through deepbrain_call and when the Key has the brain.propose permission.

Important Limits:
* No caching: All calls are live requests. Judgments may be overridden or invalidated, and caching can cause the model to retrieve outdated conclusions.
* No write permission: Unless using deepbrain_call with the brain.propose permission, the plugin does not write any data to DeepBrain.
* No client UI: This is a Host-only plugin that only provides tool registration and does not include a frontend interface.

Known Issues

In a clean profile environment (without enabling auto-install-peers), you may encounter the Cannot find package '@deepseek-ai/cordis' error.

This is because the plugin depends on @deepseek-ai/dsh-tools, which declares @deepseek-ai/cordis as a peer dependency. DSH usually provides cordis, but in certain environments it may not be installed automatically. The solution is to enable automatic installation of peer dependencies:

pnpm config set auto-install-peers true

The plugin deliberately avoids depending on its own copy of cordis to avoid fragmentation of DI container context and service identity.

Summary

dsh-deepbrain solves the problem that agent memory often lacks structured validation. It converts DeepBrain’s “judgment” capability into tools available to DSH, allowing agents to reason and cite based on validated evidence. For developers who need to handle a large volume of meeting transcripts and interview content, and who want agents with reliable memory capabilities, this is a practical tool.