dsh-plugin-reme

DeepSeek Harness plugin for ReMe long-term memory integration.

Overview

This plugin integrates ReMe's file-based long-term memory system into DeepSeek Harness, giving the agent six model-facing tools to search, read, write, and consolidate durable knowledge stored as Markdown files.

Tools Provided

Tool	Description
reme_search	Search ReMe memory by keyword/query (BM25)
reme_read	Read a memory file by workspace path
reme_write	Write a new memory file with frontmatter
reme_save_memory	Save conversation turns into daily memory cards
reme_proactive	Read proactive interest topics from auto_dream
reme_dream	Consolidate daily notes into long-term digest memory

Prerequisites

1. DeepSeek Harness installed (dsh CLI available)
2. ReMe installed (pip install "reme-ai[core]")
3. pnpm available (for DSH plugin management)

Installation

From Local Path (Development)

# Navigate to the plugin directory
cd /path/to/dsh-plugin-reme

# Add the plugin to your DSH profile
dsh plugin --profile web add ./dsh-plugin-reme

From Git Repository

dsh plugin --profile web add git+https://github.com/yourusername/dsh-plugin-reme.git

From npm Registry (When Published)

dsh plugin --profile web add dsh-plugin-reme

Configuration

The plugin reads its configuration from the cordis.patch.yml patch entry. Override defaults by adding a patch in your profile's cordis.patch.yml:

- id: tool-reme
  config:
    host: '127.0.0.1'      # ReMe service host
    port: 2333             # ReMe service port
    searchLimit: 5         # Default search result limit
    timeoutMs: 60000       # Tool call timeout in milliseconds

Usage

1. 安装插件

dsh plugin --profile web add ./dsh-plugin-reme

2. 启动 ReMe 服务

插件只是 HTTP 客户端，需要 ReMe 服务在运行。

基础启动（只有 BM25 关键词搜索）：

reme start

启用向量检索 + LLM（推荐）：

reme start \
  "components.as_embedding.default.backend=openai" \
  "components.as_embedding.default.model=你的embedding模型" \
  "components.as_embedding.default.dimensions=维度" \
  "components.as_embedding.default.credential.api_key=sk-placeholder" \
  "components.as_embedding.default.credential.base_url=http://你的embedding端点/v1" \
  "components.embedding_store.default.backend=local" \
  "components.embedding_store.default.as_embedding=default" \
  "components.file_store.default.embedding_store=default"

⚠️ 不要用 config=file.yaml 覆盖整个 components 段，这会导致内置组件（file_graph、keyword_index）丢失，启动失败。

首次启用 embedding 需要重建索引：

reme reindex

3. 验证

reme search query="测试" limit=3
# counts 中显示 "hybrid": true 说明向量检索生效

Example Agent Workflows

1. Search before answering: The agent searches ReMe for relevant prior context before answering questions about project history or user preferences.

2. Save after useful conversations: After a productive exchange, the agent calls reme_save_memory to distill key facts into daily memory cards.

3. Periodic consolidation: The agent runs reme_dream to merge daily notes into long-term digest knowledge nodes.

4. Proactive recall: The agent checks reme_proactive to discover interest topics it should bring up.

Project Structure

dsh-plugin-reme/
├── package.json              # Package manifest with dsh.bundle declaration
├── cordis.patch.yml          # Cordis loader patch (entry registration)
├── reme-start-commands.md  # ReMe 启动参数参考（CLI）
├── lib/
│   ├── index.js              # Main plugin entry (tool registrations)
│   └── constants.js          # Shared constants
└── skills/
    └── reme_memory/
        └── SKILL.md          # Agent skill documentation

How It Works

1. Cordis Patch: The cordis.patch.yml declares an entry that loads the dsh-plugin-reme module
2. Tool Registration: On activation, the module registers 6 tools with ctx.tools.register() and adds system prompt guidance via ctx.systemPrompt.section()
3. HTTP Client: Tool executions call the ReMe HTTP API (POST /{action}) on the configured host:port
4. Streaming Support: The client handles both JSON and SSE streaming responses from ReMe

License

MIT