跨会话记忆插件 — 让 AI Agent 拥有持久记忆和全局身份
Cross-session memory plugin for DeepSeek Harness (DSH)

---

English | 中文

Overview

dsh-memory-connect is a cross-session memory sharing plugin for DeepSeek Harness. It automatically extracts, stores, and recalls memories across sessions, giving your AI agent persistent, intelligent memory with context explosion prevention and global Soul identity.

Zero-config — works out of the box with SQLite FTS5 and DSH's built-in LLM.

🚨 v0.4.0 — The "it actually works now" release

v0.3.0 registered the service but never instantiated it (Cordis lazily constructs services — passing the class to ctx.provide() means the constructor never runs), and the "recall" feature wrote a computed context to a field no one ever read. In practice the plugin did nothing.

v0.4.0 fixes the activation chain and wires recall into the system prompt: a systemPrompt.context provider injects ## Related Memories from Previous Sessions on every turn. Zero-config runnable: a bare dsh plugin add used to crash on config.openAt (Cordis passes undefined config when the patch has no config: block) — apply() now fills in full defaults (DB at ~/.dsh/memory.db, openAt: startup). Verified end-to-end on dsh v0.1.1-rc.2 / Node 24: tell session A "my cat is named Mimi", start a fresh session B and ask — the model answers correctly from memory.

See CHANGELOG.md for the full breakdown.

Features

Feature	Description
🧠 Global Soul	Persistent identity across all workspaces via ~/.dsh/soul.md
🔍 Auto Extraction	Extracts facts, preferences, decisions, and context from conversations
🧠 Cross-Session Recall	FTS5 keyword recall injected into the system prompt every turn (recency fallback when no query text is available)
🛡️ Context Explosion Prevention	Token budget management prevents context window overflow
⏰ Scheduled Maintenance	Automatic periodic decay and consolidation via built-in scheduler
🤖 LLM Consolidation	Intelligent memory merging using DSH's built-in ctx.llm (zero-config)
📉 Memory Decay	Old, unused memories naturally fade; frequently accessed ones persist
🎯 Smart Prioritization	Memory selection based on relevance × recency × frequency
🗜️ Memory Compression	Automatic compression when approaching token limits
🧭 Temporal Context Graph	Every memory carries valid_from/valid_until; corrections are append-only via reviseMemory() (soft-supersede + successor link), never destructive. History stays queryable; recall only sees current truth.
🛡️ Trust Model	Recalled history is injected as an untrusted reference (explicit warning; current instruction always wins) — prevents memory poisoning and prompt conflicts
📝 Turn-End Summarization	turn/end events auto-generate lightweight summary memories preserving the conversation arc
🔎 Semantic Recall	Optional local embedding (BGE-small-zh-v1.5) fuses with FTS5 via RRF — finds memories by meaning even with zero keyword overlap. Opt-in (embeddingEnabled: true).

🧠 Global Soul (Identity)

The Soul feature provides a persistent identity that follows you across all workspaces.

How it works

1. Create ~/.dsh/soul.md with your identity information
2. The plugin automatically loads and injects it into every session
3. Your preferences, tech stack, and coding style are always available

Example Soul file

# 🧠 Soul — Global Identity

## 👤 Identity
- GitHub: your-username
- Role: Developer/Designer/Product Manager

## 💻 Tech Stack
- TypeScript, React, Node.js, DSH/Cordis

## 🎨 Coding Style
- Functional programming
- ES Modules
- Zero-config preferred

## ⚠️ Preferences
- No class components
- No redundant comments

Installation

✅ Published on npm — install from npm registry (v0.6.1):

# 方式 A：dsh plugin add（推荐，自动装入 profile）
dsh plugin --profile <你的profile> add @asherliner/dsh-memory-connect

# 方式 B：npm 直接安装
npm install @asherliner/dsh-memory-connect

# 方式 C：本地源码 link（开发/调试用）
git clone https://github.com/Asher-2000/dsh-memory-connect.git
cd dsh-memory-connect
dsh plugin --profile <你的profile> add link:/path/to/dsh-memory-connect

Add to your DSH composition:

# agent.cordis.yml
- id: memory
  name: '@asherliner/dsh-memory-connect'
  config:
    path: ~/.dsh/memory.db
    enableSoul: true  # Enable Soul injection (default: true)

Requirement: the plugin injects systemPrompt (used for the cross-session recall context provider). Any profile that loads this plugin needs the systemPrompt service available (standard in the dsh web/headless profiles).

Configuration

Basic Options

Option	Default	Description
path	(required)	Path to SQLite memory database
openAt	startup	When to open: startup, first-query, never
maxRecallCount	10	Max memories to recall per session
decayRate	0.02	Decay constant (higher = faster decay)
journalMode	wal	SQLite journal mode

Scheduler Options

Option	Default	Description
schedulerEnabled	true	Enable periodic maintenance
schedulerDecayIntervalMs	3600000	Decay interval (ms), default 1h
schedulerConsolidateIntervalMs	21600000	Consolidation interval (ms), default 6h

Context Explosion Prevention Options

Option	Default	Description
maxContextTokens	4000	Max tokens for memory context injection
smartPrioritization	true	Enable smart memory prioritization
enableCompression	true	Enable memory compression

Soul Options

Option	Default	Description
enableSoul	true	Enable global Soul injection
soulPath	~/.dsh/soul.md	Custom Soul file path

Semantic Embedding Options (optional)

Option	Default	Description
embeddingEnabled	false	Enable semantic (vector) recall via local embedding server
embeddingUrl	http://127.0.0.1:8765	Embedding server base URL
embeddingModel	BAAI/bge-small-zh-v1.5	Model name (must match the server)
embeddingWeight	0.7	RRF fusion weight for semantic results (0–1)

To use semantic recall, start the bundled embedding server first:

# one-time: install the Python model
pip install sentence-transformers

# start the server (keeps the model resident)
python3 node_modules/@asherliner/dsh-memory-connect/scripts/embed_server.py --port 8765

Then enable it in the plugin config (embeddingEnabled: true). The plugin degrades gracefully to keyword-only recall if the server is unreachable.

Full configuration example:

- id: memory
  name: '@asherliner/dsh-memory-connect'
  config:
    path: ~/.dsh/memory.db
    openAt: startup
    maxRecallCount: 10
    decayRate: 0.02
    journalMode: wal
    schedulerEnabled: true
    maxContextTokens: 4000
    smartPrioritization: true
    enableCompression: true
    enableSoul: true

API

Search Memories

const memories = await ctx.crossSessionMemory.searchMemories({
  query: 'TypeScript configuration',
  types: ['fact', 'decision'],
  limit: 5,
})

Recall for Session

const memories = await ctx.crossSessionMemory.recallForSession(
  'session-123',
  'Setting up a new React project',
  10
)

Synchronous Recall (for system-prompt providers)

// v0.4.0+ — sync API for prompt-context providers (node:sqlite is synchronous,
// so no async needed). Returns a formatted markdown block or ''.
const block = ctx.crossSessionMemory.recallSync('session-123', 'React project')

Latest User Text (from dsh session logs)

// v0.4.0+ — extracts the most recent user message text from a dsh Session
// (handles {event: ...} wrappers, agent/inbox/spliced, and bare events).
const query = ctx.crossSessionMemory.currentUserText(agent.session)

Store Memory

await ctx.crossSessionMemory.storeMemory({
  type: 'preference',
  content: 'User prefers functional programming style',
  sessionId: 'session-123',
  tags: ['coding-style', 'preference'],
})

Manual Maintenance

// Trigger a full maintenance cycle (decay + consolidation)
const result = await ctx.crossSessionMemory.triggerMaintenance()

// Or run individually
await ctx.crossSessionMemory.runDecay()
await ctx.crossSessionMemory.consolidate()

Context Explosion Prevention

How It Works

1. Token Counting — Estimates tokens for English, Chinese, and mixed text
2. Smart Prioritization — Ranks memories by: relevance × 50% + recency × 30% + frequency × 20%
3. Budget Management — Enforces maxContextTokens limit (default: 4000)
4. Memory Compression — Automatically truncates or summarizes when approaching limits

Output Example

## 🧠 Global Identity (Soul)

[Your Soul content here]

---

## Related Memories from Previous Sessions

- [preference] User prefers TypeScript
- [decision] Chose PostgreSQL over MySQL

> 💾 Memory: 2/10 memories + Soul | 250/4000 tokens

Memory Types

Type	Description	Example
fact	Objective information	"Project uses TypeScript 5.3"
preference	User preferences	"Prefers functional components"
context	Project context	"E-commerce platform migration"
decision	Decisions made	"Chose PostgreSQL over MySQL"
skill	Learned patterns	"How to configure ESLint"

Development

git clone https://github.com/Asher-2000/dsh-memory-connect.git
cd dsh-memory-connect
npm install
npm test

License

MIT

---

中文

概述

dsh-memory-connect 是 DeepSeek Harness 的跨会话记忆共享插件。它自动从对话中提取、存储和检索记忆，让 AI Agent 拥有持久化的智能记忆能力，并防止上下文爆炸和全局身份。

零配置 — 基于 SQLite FTS5 和 DSH 内置 LLM，开箱即用。

🚨 v0.4.0 — "这次真的能跑"版本

v0.3.0 只注册了服务但从未实例化（Cordis 懒加载机制：把类传给 ctx.provide() 导致构造函数永远不执行），且"召回"功能把计算好的上下文写进了一个无人读取的字段——实际运行中插件什么都不做。

v0.4.0 修复了启动链路，并把召回真正接入了系统提示词：通过 systemPrompt.context 提供者，每轮对话都注入 ## Related Memories from Previous Sessions。零配置即可运行：此前仅执行 dsh plugin add（patch 无 config: 块时 Cordis 传入 undefined 配置）会在 config.openAt 崩溃——现在 apply() 自动填充完整默认值（DB ~/.dsh/memory.db、openAt: startup）。已在 dsh v0.1.1-rc.2 / Node 24 上端到端验证：会话 A 说"我的猫叫咪咪"，开新会话 B 问它，模型能正确从记忆中回答。

详见 CHANGELOG.md。

核心功能

功能	说明
🧠 全局身份 (Soul)	通过 ~/.dsh/soul.md 跨所有工作区持久化身份
🔍 自动提取	从对话中提取事实、偏好、决策和上下文
🧠 跨会话召回	FTS5 关键词召回，每轮注入系统提示词（无查询文本时按最近记忆兜底）
🛡️ 上下文爆炸防护	Token 预算管理，防止上下文窗口溢出
⏰ 定时维护	内置调度器自动执行衰减和整合
🤖 LLM 整合	使用 DSH 内置 LLM 智能合并相似记忆
📉 记忆衰减	旧的、不常用的记忆自然消退
🎯 智能优先级	基于相关性 × 时间 × 频率的记忆排序
🗜️ 记忆压缩	接近 token 限制时自动压缩
🧭 时态上下文图谱	每条记忆带 valid_from/valid_until；修正通过 reviseMemory() 追加而非覆盖（软废弃旧记忆 + 新记忆链接），历史可回溯，召回只见当前有效真相
🛡️ 信任模型	召回的历史作为不可信参考注入（显式警告；当前指令绝对优先）— 防止记忆投毒和提示词冲突
📝 轮末自动摘要	turn/end 事件自动生成轻量 summary 记忆，保留对话脉络

🧠 全局身份 (Soul)

Soul 功能提供跨所有工作区的持久化身份。

工作原理

1. 创建 ~/.dsh/soul.md 包含你的身份信息
2. 插件自动加载并注入到每个会话
3. 你的偏好、技术栈和编码风格始终可用

Soul 文件示例

# 🧠 Soul — 全局身份

## 👤 身份
- GitHub: your-username
- 角色: 开发者/设计师/产品经理

## 💻 技术栈
- TypeScript, React, Node.js, DSH/Cordis

## 🎨 编码风格
- 函数式编程
- ES Modules
- 零配置优先

## ⚠️ 偏好
- 不用 class 组件
- 不写冗余注释

快速开始

✅ 已发布到 npm — 从 npm registry 安装 (v0.6.1)：

# 方式 A：dsh plugin add（推荐，自动装入 profile）
dsh plugin --profile <你的profile> add @asherliner/dsh-memory-connect

# 方式 B：npm 直接安装
npm install @asherliner/dsh-memory-connect

# 方式 C：本地源码 link（开发/调试用）
git clone https://github.com/Asher-2000/dsh-memory-connect.git
dsh plugin --profile <你的profile> add link:/path/to/dsh-memory-connect

添加到 DSH 配置：

# agent.cordis.yml
- id: memory
  name: '@asherliner/dsh-memory-connect'
  config:
    path: ~/.dsh/memory.db
    enableSoul: true

依赖说明：插件注入 systemPrompt 服务（用于跨会话召回 context 提供者）。所在 profile 需要提供 systemPrompt（dsh web/headless profile 默认都有）。

许可证

MIT