One-Sentence Pitch

Give DeepSeek Harness cross-session persistent memory — let the AI keep what it learned about your preferences, projects, decisions in this conversation, and automatically bring them back into the system prompt for new sessions.

Core Capabilities

• Auto-write and deduplication merge: The model stores memories via memory_save tool; entries with the same title merge automatically with version history preserved (content_history)
• Auto-inject: At session start, high-importance memories are injected into the system prompt so the AI sees what you said before
• Background autoDream auto-organization: After memory count or word count thresholds are met, async LLM arbitration triggers (keep / merge / archive / conflict / update); the memory bank gets more refined over time
• Editable Markdown mirror: Preferences, projects, decisions, history, and overview each land in separate .md files, with human edits taking priority for write-back
• Fully optional semantic search: Beyond keywords, add BM25 + vector retrieval + cross-encoder reranking; three-way retrieval fusion with adaptive thresholds
• Session summary + short-term hot memory: Session end automatically extracts key points; the last 5 conversation turns in current session are injected as rolling hot memory blocks
• Entity memory + Wiki-Link + Tag system: Extract named entities and timeline attributes from memories; supports [[bidirectional links]] and #tags (disabled by default, enable as needed)

Technical Implementation

• Language: JavaScript (Node.js ESM, type: module)
• Key Dependencies: @deepseek-ai/cordis (DI container), @deepseek-ai/dsh-tools (model tool registration), @deepseek-ai/schemastery (config Schema), @deepseek-ai/transformers (optional, local vector/reranking)
• Architecture Pattern: Cordis apply(ctx, config) injection into several host hooks: tools / systemPrompt / webServer / llm / commands; default config written to DSH profile via cordis.patch.yml bundle for zero-config startup; background maintenance via two pipelines: createDreamScheduler threshold-based scheduling + createSleepScheduler idle-based scheduling
• Entry Point: dsh-mneme/lib/index.js (published artifact), development source at dsh-mneme/src/index.js

Use Cases

This plugin solves the problem when you want DSH to remember your project background, coding habits, and past decisions like a long-term coworker — instead of re-explaining everything every new session. It's suited for medium-to-long-term users: those who work with AI for several hours daily and need a retrievable knowledge base accumulating over time. It's not suitable for one-off Q&A or chat-and-go scenarios. With semantic search enabled, the AI can also recall things with natural language like "that thing we talked about last time," allowing you to continue projects across days without restating context.

Prerequisites & Compatibility

Installation

dsh plugin --profile web add github:modusensus/dsh-mneme

Configuration Options

Default behavior works out of the box, all listed in schema with no hard requirements. Below are the most commonly adjusted items; modify in DSH Settings → Memory Settings as needed, or override in ~/.dsh/profiles/web/cordis.patch.yml with id: dsh-mneme section.

FAQ

Q: I installed the plugin but have no memory available. Do I need to feed it conversation content first?

A: No need. When you open a new session, the AI will see existing memories; new facts are written by the AI itself via the memory_save tool — no manual calls required.

Q: Where is data stored? Will it sync to the cloud?

A: Only on the machine running DSH, at path ~/.dsh/memory/. SQLite main database and Markdown mirror are together; no remote upload. Just copy the directory to back up.

Q: If I delete the conversation window, does the memory go away too?

A: By default no — memories and conversation windows are separate; deleting a session doesn't affect saved memories. With sessionLifecycleEnabled enabled, deleting a session becomes soft-hide (you can still see and recover via memory_list with include_archived).

Q: How do I let the AI search with natural language like "that thing we talked about last time"?

A: Go to Settings → Memory → Toolbar and switch to "Semantic"; this enables vector retrieval. Background is keyword + BM25 + vector three-way fusion with query-length-adaptive thresholds.

Q: My embedding service is unreachable — will vectors cause errors?

A: No. Embedding failures automatically fall back to keyword search; local ONNX model download failures just log warnings and degrade to keyword search — main flow works normally.

Q: Why does a deleted memory still show up in search?

A: You might be using memory_forget or memory_archive — these suppress injection and archive respectively, both recoverable. To physically delete, explicitly use memory_delete, either by id or by query description.

Q: How do I completely disable this plugin?

A: Remove the bundle from DSH profile: dsh plugin --profile web remove @modusensus/dsh-mneme. Data stays on disk; reinstall to restore.

Learning Curve

Beginner — default config works out of the box; most users don't need to change any settings. Only when preparing to enable optional features like vector search / Sleep Mode / entity extraction do you need to check the config options.

Known Issues & Limitations

• Reasoning models (e.g., deepseek-v4-flash / DeepSeek-R1) will burn the entire token budget on reasoning with autoDream default parameters, resulting in empty body (common log: no json array in llm output). Fix: explicitly set both dreamReasoningEffort and sleepReasoningEffort to off; these default to none to preserve existing behavior.
• Rerank reranking is disabled by default (rerankEnabled=false) because initialization loads onnxruntime-node; to use it, you must explicitly enable it and accept model download + first-inference latency.
• Database is single-file SQLite (node:sqlite synchronous API), doesn't support concurrent multi-process writes; DSH runs single-process hosting, multi-instance sharing will conflict.
• LLM routing goes through agentDefaultModel, no independent key: by default consumes main conversation token budget; when budget is tight, manually set dreamProvider / dreamModel / summarizeProvider to point to lightweight models.
• Advanced features (entity extraction, Wiki-Link, Tag weighting, Sleep, Session lifecycle) are disabled by default — conservative enabling to avoid breaking existing behavior; consult config docs before first use.

dsh-mneme

中文 | English

---

🇨🇳 dsh-mneme（中文）

记忆基因 · 让记忆自我进化 —— 从文本仓库到结构化知识库，记忆不再只是存储，而是会生长。

dsh-mneme 是一个 DeepSeek Harness (DSH) 插件，为 Agent 提供持久的跨会话记忆能力。Mneme（Μνήμη）——希腊记忆女神 Mnemosyne 之名，掌管记忆与梦境，正如 autoDream 在后台巩固记忆。

它能解决什么问题

每次新开对话，AI 都像第一次认识你？

dsh-mneme 给 DeepSeek Harness 装上跨会话记忆。 你聊过的项目、提过的偏好、做过的决定，AI 都记得——即使关掉了窗口，下次打开还在。

三个典型场景

核心特性

• 跨会话记忆 — 对话中 AI 自动记录关键信息，新会话自动注入相关记忆
• 自动整理 — 后台自动去重、合并、归档，记忆库越用越精炼
• 删对话 ≠ 删记忆 — 删除聊天窗口不会丢掉已保存的记忆（可配置）
• 你的数据你做主 — 所有记忆以 Markdown 格式存于本地，随时打开查看和编辑
• 完全离线 — 默认无需 API Key，所有处理在本地完成

安装

# 安装插件（自动注册 bundle 层）
dsh plugin --profile web add @modusensus/dsh-mneme
dsh web

需要 Node 24+（node:sqlite）。完整安装 / 配置 / 架构见 插件文档。

快速配置（可选）

装完即用，以下按需开启：

在 DSH 设置面板 → 记忆库设置 中修改。完整配置说明见 配置文档。

隐私承诺

• 数据只存在你的电脑本地，不上传任何服务器
• 记忆是 Markdown 文件，人类可读、可手工编辑
• 默认零网络依赖，不需要 API Key
• 无遥测、无分析、无远程日志

文档

🗺️ 路线图

🧬 基因（v0.3.0）→ 🛡️ 审计加固（v0.3.6–0.3.9）→ 💤 睡眠维护（v0.4.0）→ 🕸️ 召回融合与图谱（v0.5.0）→ ✨ 自进化（v0.6.0）→ 🔮 兴趣漂移 + 跨 workspace（v0.7.0+）

🧪 本地开发

cd dsh-mneme
npm install
npm test          # 628 个测试
npm run stress    # 三轴线压测
npm run sync      # src → lib 同步

📜 License

MIT

---

🇬🇧 dsh-mneme (English)

Memory Genome · Let memory evolve — from text warehouse to structured knowledge base. Memory is no longer just stored; it grows.

dsh-mneme is a DeepSeek Harness (DSH) plugin providing persistent cross-session memory. Mneme (Μνήμη) — named after Mnemosyne, the Greek goddess of memory and dreams, mirroring how autoDream consolidates memories in the background.

What it does

Every time you start a new chat, the AI acts like it's never met you?

dsh-mneme gives DeepSeek Harness cross-session memory. Projects you've discussed, preferences you've mentioned, decisions you've made — the AI remembers them even after you close the window.

Three typical scenarios

Core Features

• Cross-session memory — AI automatically records key info during chats; new sessions inject relevant memories
• Auto-consolidation — Background deduplication, merging, and archival; the memory store refines itself over time
• Delete session ≠ delete memories — Closing a chat window doesn't lose saved memories (configurable)
• Your data, your control — All memories stored as local Markdown files, human-readable and editable
• Fully offline — Zero API Key needed by default; all processing happens locally

Install

dsh plugin --profile web add @modusensus/dsh-mneme
dsh web

Requires Node 24+ (node:sqlite). Full install / config / architecture docs in the plugin README.

Quick Config (Optional)

Works out of the box. Enable these as needed:

Change in DSH Settings Panel → Memory Settings. Full config docs here.

Privacy

• Data stays on your machine only, never uploaded
• Memories are Markdown files, human-readable and editable
• Zero network dependency by default, no API Key required
• No telemetry, no analytics, no remote logging

Docs

🗺️ Roadmap

🧬 Gene (v0.3.0) → 🛡️ Audit hardening (v0.3.6–0.3.9) → 💤 Sleep maintenance (v0.4.0) → 🕸️ Recall fusion & graph (v0.5.0) → ✨ Self-evolve (v0.6.0) → 🔮 Interest drift + cross-workspace (v0.7.0+)

🧪 Local Development

cd dsh-mneme
npm install
npm test          # 628 tests
npm run stress    # three-axis stress test
npm run sync      # src → lib sync

📜 License

MIT