DeepSeek Harness Plugin Hub

Publish and manage complete Harness Profiles. Discover Plugins for your next setup.

Explore

PluginsPresetsDocsNews

Community

Publish a pluginContactReport an issue

Resources

Plugin Hub on GitHubDeepSeek HarnessSystem statusPrivacy notice
© 2026 DeepSeek Harness Plugin HubPowered byPaxTech

Independent and unofficial. Not affiliated with, authorized by, or endorsed by DeepSeek.

Local Vector Memory — DSH Plugin for DeepSeek Harness
DeepSeek Harness Plugin Hub
ProfilesPluginsCategoriesNewsDocsSign inManage Profiles
ProfilesPluginsCategoriesNewsDocsSign in
← Plugins
L

dsh-local-vector-memory

Local Vector Memory

DSH local vector memory plugin: local embedding vectorization + SQLite storage; automatic recall with four-signal hybrid retrieval (vector + BM25 + temporal freshness + tag matching); automatic extraction via cloud-based deepseek-v4-pro; supports replacement chains, pinning, soft-delete recycle bin,

The plugin will be installed here. Keep web if you are unsure.

npx -y @deepseek-ai/dsh plugin --profile web add github:liangxiaobing520/dsh-local-vector-memory#56d5e4cbbb62eed18b478f689946daa253756769
READMECompatibilityVersions

Description

DSH local vector memory plugin: local embedding vectorization + SQLite storage; automatic recall with four-signal hybrid retrieval (vector + BM25 + temporal freshness + tag matching); automatic extraction via cloud-based deepseek-v4-pro; supports replacement chains, pinning, soft-delete recycle bin, memory_merge merging, memory_health reports, and scheduled backups.

Compatibility and provenance

Local Vector Memory is published as dsh-local-vector-memory and currently resolves to version 0.3.3. The Hub verifies its manifest and preserves the exact installation source for reproducible installs.

DSH compatibility
*
Runtime surfaces
any
Release source
github
Registry updated
9/14/2026

Versions

0.3.3stable
9/14/2026
0.2.1stable
8/31/2026
0.1.0stable
8/29/2026

Related plugins

Loading related plugins…

Latest
0.3.3
DSH
*
HMR
Process restart
Tree shaking
Safe tree shaking not declared
Unpacked size
Unavailable
Files
Unavailable
Surface
any
License
MIT
Source
github
GitHub
★ 1
Weekly downloads
0
Last push
9/14/2026
View source ↗
README badge

Click the badge to copy Markdown for your README.

Do you maintain this Plugin?Claim benefit · Priority security scan

Verify the GitHub repository declared in package.json to manage this listing. After you claim it, Hub will prioritize a security scan of the current version and publish the result when it passes.

Claim this Plugin →
Report an issue

Related plugins

More verified plugins in memory-context.

Mnemondsh-mnemonComposable three-tier memory control plane for DeepSeek Harness: persistent runtime context, searchable project documents, pluggable long-term memory, guarded strategies, WebUI, and headless tools.Memory@furongjun1999/dsh-memoryLingshu (Lingshu·líng shū) DeepSeek Harness plugin: a complete brain—long-term memory/knowledge flywheel/self-awareness/recursive reflection integrated with DSH, with conversations automatically distilled into the md_cg cognitive graph (md documents)Reme@agentscope-ai/remeReMe client and memory integrations for TypeScript agentsRewind Plugindsh-rewind-pluginIn-window conversation rewind with workspace file restore · 同窗口内对话回退并可还原工作区文件

README

dsh-local-vector-memory

本地向量记忆插件 for DeepSeek Harness(DSH)。本地 embedding 向量化 + SQLite 单文件存储,对话自动召回注入;会话结束由 DeepSeek 云端 deepseek-v4-pro 自动提取。

A local-first vector memory plugin for DeepSeek Harness: local embeddings, SQLite storage, automatic recall injection, and DeepSeek cloud v4-pro extraction at session flush.

特性

  • 写入三条路:手动 memory_add;用户消息命中记忆线索(记住/以后/偏好/约定/不要…)时毫秒级自动捕获;DeepSeek 云端 deepseek-v4-pro 在会话结束时从整段对话提取记忆(autoExtract,默认开)
  • 防矛盾记忆:写入时检测与已有记忆的冲突/过时(相似度 ≥0.86 但未达完全重复),提示改用 memory_update 更新旧记忆而不是新增重复条目
  • 回收站:memory_forget 默认软删除,memory_restore 可恢复;purge=true 才永久删除
  • 在线备份:memory_backup 用 SQLite VACUUM INTO 生成一致性快照(安全于手工复制 WAL 库),默认保留 5 份
  • 自动召回:每次 agent 推理前检索 top-K 相关记忆注入上下文(<local-memory> 标签),会话内 LRU 去重
  • 批量向量化:提取/重建索引一次 HTTP 批量请求,失败自动逐条回退
  • 关键词兜底:embedding 服务不可用时退化为中英文关键词匹配,写入照常
  • 本地存储:向量库是单个 SQLite 文件(node:sqlite,Node ≥22.5);embedding 全本地,提取走 DeepSeek 云端

需求

  • DSH(DeepSeek Harness),web profile
  • Node.js ≥ 22.5(内置 node:sqlite)
  • OpenAI 兼容的本地 embedding 服务(见下文,推荐 Qwen3-Embedding-0.6B + llama-server)
  • DeepSeek 官方云端 API(deepseek-v4-pro)用于提取(可选;默认走线索自动捕获)
  • ⚠️ 不要把 API key 写进任何会被提交的文件。key 只放在本机,支持三种写法:
    extractionApiKey: "${DEEPSEEK_API_KEY}"        # 1. 环境变量
    extractionApiKey: "dsh:DEEPSEEK_API_KEY"       # 2. DSH 凭据库(~/.dsh/.credentials.yaml)
    extractionApiKey: "local-no-auth"              # 3. 明文占位(仅限免认证的本地服务)
    
    仓库内的 cordis.patch.yml 只是 bundle 挂载声明,不含任何凭据;profiles/<profile>/cordis.patch.yml 用于覆盖配置,请勿在其中写明文 key(会留在磁盘上,也可能被 UI 回显)。

安装

dsh plugin --profile web add dsh-local-vector-memory

插件声明 dsh.bundle.patch,安装后 dsh.profile.bundles 自动追加,无需手改 package.json。重启 DSH Web 生效。

工具(12 个)

工具作用
memory_add写入一条长期记忆(用户说"记住…"时;自动去重 + 冲突检测)
memory_search向量检索(embedding 挂了自动关键词兜底;支持 tag 过滤)
memory_list浏览记忆,支持 tag/source 过滤;includeDeleted=true 看回收站
memory_update按 id 更新文本/标签(自动重向量化),纠正过时/冲突记忆
memory_forget按 id 删除,默认软删除(可恢复);purge=true 永久删除
memory_restore恢复被软删除的记忆(回收站)
memory_backupSQLite 在线备份到 ~/.dsh/backups/memory,自动保留最近 5 份
memory_stats库状态 + 软删除数 + 服务地址
memory_reindex给未向量化条目补向量(批量)
memory_extract调用提取模型从一段文本提炼记忆入库(仅短文本,建议 ≤2000 字)
memory_merge合并重复/同主题记忆:保留一条,其余软删除并标记"已被取代"
memory_health健康报告:条数/软删/未向量化/置顶分布 + 重复对检测与合并建议

默认配置

全部默认值在 lib/config.mjs。覆盖方式:在 ~/.dsh/profiles/web/cordis.patch.yml 写:

- id: local-vector-memory
  config:
    autoRecall: true
    recallTopK: 6
    recallMinScore: 0.55
    autoExtract: false   # 可选实验项:若提取模型是长思考链模型,长文本会耗尽 max_tokens,谨慎开启
    autoCapture: true    # 用户消息命中"记住/以后/偏好/约定/不要"等线索时自动入库
    skipDuplicates: true # 相同/高度相似记忆跳过
    softDelete: true     # memory_forget 默认软删除(回收站)
    backupDir: ~/.dsh/backups/memory
    backupKeep: 5
    conflictScore: 0.86  # 相似度 >= 此值(未达完全重复)提示冲突、建议 memory_update
    recallDedupeLru: 16  # 会话内召回去重 LRU 上限
    embeddingEndpoint: auto      # auto = WSL 网关 + embeddingPort
    embeddingPort: 8081
    extractionEndpoint: auto     # auto = WSL 网关 + extractionPort
    extractionPort: 8080

注意 DSH patch 是整段替换,不是深合并;上例只写了要改的字段,其余回到默认值。

配套 embedding 服务(独立于插件)

模型下载(一次):

modelscope download --model Qwen/Qwen3-Embedding-0.6B-GGUF \
  Qwen3-Embedding-0.6B-Q8_0.gguf --local-dir /path/to/models

Windows 启动脚本(路径按你的实际安装位置替换):

# <LLAMA_DIR> = llama.cpp 可执行文件目录;<MODEL_DIR> = GGUF 模型目录
Start-Process -FilePath '<LLAMA_DIR>\llama-server.exe' -ArgumentList @(
  '-m','<MODEL_DIR>\Qwen3-Embedding-0.6B-Q8_0.gguf',
  '--embeddings','--pooling','mean','-c','8192',
  '--host','0.0.0.0','--port','8081','-ngl','0'
) -WindowStyle Hidden -RedirectStandardOutput '<LLAMA_DIR>\embedding-server.log' -RedirectStandardError '<LLAMA_DIR>\embedding-server.err.log' -PassThru

验证:

# <GATEWAY> = embedding 服务所在主机地址(同机用 127.0.0.1;WSL 访问 Windows 宿主用默认网关)
curl http://<GATEWAY>:8081/v1/embeddings \
  -H 'Content-Type: application/json' \
  -d '{"model":"local-embedding","input":"测试"}'

数据流

用户消息 ──记忆线索命中?──▶ embedding(8081) ──▶ SQLite 向量库
memory_add ─────────────────────────────▶ embedding(8081) ──▶ SQLite 向量库
memory_extract ──▶ 本地 9B(8080,仅短文本) ──▶ 批量 embedding ──▶ SQLite 向量库
                                                       │
                 agent/pre-step 自动召回 ◀──────────────┘

故障排查

  • memory_stats 显示未向量化很多:检查 8081 服务,然后 memory_reindex。
  • 自动提取没发生:autoExtract 默认关闭。若 extractionModel 指向长思考链模型,长文本提取会耗尽 max_tokens(实测),故默认改为线索自动捕获;短文本可手动 memory_extract。
  • DSH 日志过滤:[dsh-local-vector-memory]。
  • WSL 重启后网关变化:插件自动探测默认网关;若 8081/8080 不通,先确认 Windows 侧服务在跑。

License

MIT