记忆主权，归还于你 —— 记忆不再是黑盒，而是你读得懂、改得动的 Markdown。
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。
不同于把记忆锁进数据库的插件，Mneme 把记忆写成你读得懂的 Markdown——你始终握着记忆的主权：看得见、改得动、删得掉，记忆这回事不该让 Agent 一个人说了算。
- 🧠 记忆主权：SQLite + 可人工编辑的 Markdown 镜像，双向同步——记忆透明、可审查、归你所有
- autoDream 梦境巩固：后台自动去重 / 合并 / 归档 / 冲突裁决 / 自我修正（fail-safe 校验），越用越精炼
- 反思更新（v0.2.1）：autoDream 可直接修正单条过时/错误记忆；失败追踪表记录纠正历史，为自进化积累数据
- 7 个模型工具：
memory_save/memory_search/memory_list/memory_update/memory_delete/memory_forget/memory_archive - 自动注入 + 会话摘要：新会话自动带入相关记忆，会话结束自动提炼偏好 / 决策 / 教训
- Web 记忆面板：官方设置面板内嵌，按类型浏览、全文搜索 + 语义（向量）搜索
- 用户设置 + 自定义指令：用户画像、行为规则每轮注入；注册斜杠命令
- 向量搜索：OpenAI 兼容 embeddings API，语义匹配字面不同但意思相近的记忆
完全离线的语义记忆引擎——embedding、rerank、搜索全在本地，零 API 成本：
- 本地 Embedding：三后端可选——ONNX（
Xenova/bge-small-zh-v1.5，离线）/ Ollama / OpenAI 兼容，失败自动逐级降级，最差回退关键词搜索 - Rerank 精排：
Xenova/bge-reranker-base对召回候选交叉编码精排，提升 Top-K 准确率 - autoDream 语义增强：对记忆向量聚类（
clusterMemories），自动发现主题相近 / 疑似矛盾的记忆，巩固更精准 - 搜索流水线：混合召回（关键词 + 向量）→ Rerank → Top-K
在 cordis.patch.yml 配置 embedProvider（默认 openai 保持 v0.1 行为，切到 local 即完全离线）。无需数据迁移。
# 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）。安装 / 配置 / 架构详见 插件完整文档。
dsh-mneme/ 插件本体（npm 包 @modusensus/dsh-mneme）
docs/ 设计文档与实施计划
cd dsh-mneme
npm install
npm test # 258 个测试
npm run stress # 三轴线压测（长会话检索 / 冲突仲裁 / 多 Agent 并发）
npm run sync # src → lib 同步（发布时自动执行）| 文档 | 路径 |
|---|---|
| 插件完整文档（功能 / 安装 / 配置 / 架构） | dsh-mneme/README.md |
| 语义架构 | dsh-mneme/docs/SEMANTIC.md |
| 本地模型部署指南 | dsh-mneme/docs/LOCAL_MODEL.md |
| v0.1 迁移说明 | dsh-mneme/docs/MIGRATION.md |
| 插件设计 | docs/superpowers/specs/2026-08-13-dsh-mneme-design.md |
| autoDream 设计 | docs/superpowers/specs/2026-08-13-dsh-mneme-autodream-design.md |
| 实施计划（核心插件） | docs/superpowers/plans/2026-08-13-dsh-memory.md |
| 实施计划（autoDream） | docs/superpowers/plans/2026-08-13-dsh-memory-autodream.md |
| 演进计划书（v0.2.1+） | docs/superpowers/plans/2026-08-15-dsh-mneme-reflection-update.md |
计划基于社区反馈持续演进，实施周期为估算值。
| 版本 | 核心主题 | 一句话 | 实施周期 |
|---|---|---|---|
| v0.2.1 | 反思更新 | ✅ 已完成（update 决策 + failure_memories 表，258 测试） | — |
| v0.2.2 | 流水线补全 | ✅ 已完成（专项测试 + 审查修复，258 测试） | — |
| v0.2.4 | 安全加固 | ✅ 已完成（API 鉴权 apiToken + apiKey 掩码 + timing-safe，258 测试） | — |
| v0.2.5 | 安全审计修复 | ✅ 已完成（CAS + 事务化应用 + reconcile 审计 + reranker opt-in + 归档恢复，258 测试） | — |
| v0.2.6 | 轻量结构化 | 实体注释标记（零 schema 变更，验证可行性） | 2-3 天 |
| v0.3.0 | 记忆基因 | 正式 entities/attrs/relations 三表 + 时间轴 | 1 周 |
| v0.3.1 | 系统级睡眠 | Sleep 调度器 + 分层压缩 + 模式发现 | 3-5 天 |
| v0.4.0 | 反思性成长 | 纠错双向回流 + 规则演进 + 自适应参数 | 1-2 周 |
| v0.5.0+ | 自进化记忆 | 兴趣漂移 + 跨 workspace（等 DSH 支持） | 远期 |
MIT
Memory sovereignty, returned to you — memory is no longer a black box, but Markdown you can read and edit.
dsh-mneme is a DeepSeek Harness (DSH) plugin that gives agents persistent cross-session memory. Mneme (Μνήμη) — named after Mnemosyne, the Greek goddess of memory and dreams, mirroring how autoDream consolidates memories in the background.
Unlike plugins that lock memory inside a database, Mneme writes memory as human-readable Markdown — memory sovereignty stays with you: see it, edit it, delete it. Memory shouldn't be decided by the agent alone.
- 🧠 Memory sovereignty: SQLite + human-editable Markdown mirror, two-way sync — memory is transparent, auditable, and yours
- autoDream consolidation: background dedup / merge / archive / conflict resolution / self-correction (fail-safe validation), refined with use
- Reflection update (v0.2.1): autoDream can correct stale/wrong single memories; a failure-tracking table records correction history for future self-evolution
- 7 model tools:
memory_save/memory_search/memory_list/memory_update/memory_delete/memory_forget/memory_archive - Auto-injection + session summary: relevant memories injected at session start, preferences / decisions / lessons distilled at session end
- Web memory panel: embedded in the official settings panel — browse by type, full-text + semantic (vector) search
- User settings + custom commands: user profile and behavior rules injected every turn; register slash commands
- Vector search: OpenAI-compatible embeddings API for semantic matching of differently-worded but related memories
A fully-offline semantic memory engine — embedding, rerank and search all run locally, zero API cost:
- Local embedding: three interchangeable backends — ONNX (
Xenova/bge-small-zh-v1.5, offline) / Ollama / OpenAI-compatible — degrading automatically, falling back to keyword search at worst - Rerank:
Xenova/bge-reranker-basecross-encoder re-ranking of recall candidates for sharper Top-K - autoDream semantic boost: vector clustering (
clusterMemories) surfaces topically-close or potentially conflicting memories for more precise consolidation - Search pipeline: hybrid recall (keyword + vector) → rerank → Top-K
Configure embedProvider in cordis.patch.yml (default openai keeps v0.1 behavior; switch to local for fully offline). No data migration needed.
📖 See Semantic architecture · Local model guide · v0.1 migration
# Install the plugin (auto-registers the bundle layer)
dsh plugin --profile web add @modusensus/dsh-mneme
dsh webRequires Node 24+ (
node:sqlite). Full install / config / architecture docs in the plugin README.
dsh-mneme/ plugin package (npm @modusensus/dsh-mneme)
docs/ design docs & implementation plans
cd dsh-mneme
npm install
npm test # 258 tests
npm run stress # three-axis stress test (long-session retrieval / conflict arbitration / concurrent agents)
npm run sync # src → lib sync (runs automatically on publish)| Doc | Path |
|---|---|
| Full plugin docs (features / install / config / architecture) | dsh-mneme/README.md |
| Semantic architecture | dsh-mneme/docs/SEMANTIC.md |
| Local model guide | dsh-mneme/docs/LOCAL_MODEL.md |
| v0.1 migration | dsh-mneme/docs/MIGRATION.md |
| Plugin design | docs/superpowers/specs/2026-08-13-dsh-mneme-design.md |
| autoDream design | docs/superpowers/specs/2026-08-13-dsh-mneme-autodream-design.md |
| Implementation plan (core) | docs/superpowers/plans/2026-08-13-dsh-memory.md |
| Implementation plan (autoDream) | docs/superpowers/plans/2026-08-13-dsh-memory-autodream.md |
| Evolution plan (v0.2.1+) | docs/superpowers/plans/2026-08-15-dsh-mneme-reflection-update.md |
Plans evolve with community feedback; timelines are estimates.
| Version | Core theme | One-liner | Timeline |
|---|---|---|---|
| v0.2.1 | Reflection update | ✅ Done (update decision + failure_memories, 258 tests) | — |
| v0.2.2 | Pipeline completion | ✅ Done (specialized tests + review fixes, 258 tests) | — |
| v0.2.4 | Security hardening | ✅ Done (API auth apiToken + apiKey masking + timing-safe, 258 tests) | — |
| v0.2.5 | Security audit fixes | ✅ Done (CAS + transactional apply + reconcile audit + reranker opt-in + archive restore, 258 tests) | — |
| v0.2.6 | Lightweight structure | Entity comment markers (zero schema change, feasibility check) | 2-3 days |
| v0.3.0 | Memory genome | Formal entities/attrs/relations tables + timeline | 1 week |
| v0.3.1 | System-level sleep | Sleep scheduler + tiered compression + pattern discovery | 3-5 days |
| v0.4.0 | Reflective growth | Correction feedback loop + rule evolution + adaptive params | 1-2 weeks |
| v0.5.0+ | Self-evolving memory | Interest drift + cross-workspace (when DSH supports it) | Long-term |
MIT