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.

Engram Relay — DSH Plugin for DeepSeek Harness
← Plugins
E

@dsh-external/dsh-engram-relay

Engram Relay

Large engram, small KV: an external engram adapter layer (cross-session hierarchical memory: global/project/session, AI-driven assignment decisions + causal link indexing + progressive disclosure + proactive awakening), implementing ultra-long-context memory sparse routing for DSH and coexisting wit

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

npx -y @deepseek-ai/dsh plugin --profile web add github:yjh051108/dsh-engram-relay#9583f973e074b07c87813896e16b7ba8fda7930e
READMECompatibilityVersions

Description

Large engram, small KV: an external engram adapter layer (cross-session hierarchical memory: global/project/session, AI-driven assignment decisions + causal link indexing + progressive disclosure + proactive awakening), implementing ultra-long-context memory sparse routing for DSH and coexisting with the official compact (the official compact frees KV; engram retains the details)

Compatibility and provenance

Engram Relay is published as @dsh-external/dsh-engram-relay and currently resolves to version 0.4.3. The Hub verifies its manifest and preserves the exact installation source for reproducible installs.

DSH compatibility
*
Runtime surfaces
web
Release source
github
Registry updated
9/19/2026

Versions

0.4.3stable
9/19/2026
0.4.2stable
9/19/2026
0.4.1stable
9/18/2026
Show 3 more versionsCollapse versions
0.3.74stable
8/21/2026
0.3.12stable
8/21/2026
0.2.0stable
8/21/2026

Related plugins

Loading related plugins…

Latest
0.4.3
DSH
*
HMR
Process restart
Tree shaking
Safe tree shaking not declared
Unpacked size
Unavailable
Files
Unavailable
Surface
web
License
Apache-2.0
Source
github
GitHub
★ 3
Weekly downloads
0
Last push
9/19/2026
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
DeepSeek Harness Plugin Hub
ProfilesPluginsCategoriesNewsDocsSign inManage Profiles
ProfilesPluginsCategoriesNewsDocsSign in

Related plugins

More verified plugins in memory-context.

Memory Plugin@openviking/dsh-memory-pluginOpenViking memory and context bundle for DeepSeek HarnessContextdsh-contextA DeepSeek Harness plugin for context insight and management, with context dashboard and context command, for understanding how the context is made of, and how it evolves.Weknora@wxg-prc-cpg/dsh-weknoraWeKnora knowledge retrieval tools for DeepSeek Harness (dsh): semantic search, document reading and RAG/agent answers over your own knowledge bases.Memsearch Dsh@zilliz/memsearch-dshMemSearch plugin for DeepSeek Harness: shared markdown memory across agents, with capture, pre-step context injection, memory-recall skill, and a skill-candidate review panel.

README

风识 FengShi · DSH 统一大脑

跨会话记忆图谱(agent 自组织)+ 灵枢知识校准器(白箱验证)+ 浅思维自动注入 + 自动补卡/补簇 + 跨端联动(知+忆)。 让 DSH 的 agent 拥有「记得住、懂得多、不硬答、当场学、学得快」的外置大脑。 百查询出招正确率 ~90%;闲聊/新域零污染(诚实边界);使用者自己补卡/补簇当场生效。

是什么

风识是一个 DSH 插件,把记忆与知识融合为一张自组织语义图:

一体两器官(r60 定版,详见 AGENTS.md 同名节):知识=去情景可共享(学科卡),经验=带情景第一人称(节点/因果/确认制);融合在接缝不在合并——verify 四态带 📎 相关记忆、respond 命中亦回头找经验、双不会共学一环、命名空间互不吞并(合并是幻觉,互见才是融合)。

统一自适应语义图(agent 建边 · 强化遗忘 · 软簇)
 ├─ 存储层:节点=记忆/概念(蒸馏+agent 织网),跨会话分层 global/project/session
 ├─ 召回层:哈希+纯算法语义匹配(SemanticScorer 三通道,零模型)
 ├─ 浅思维:图上算子(条件/验证/边界)→ 每轮自动注入 3 行
 └─ 深挖层:15 个工具(recall/store/open/link/verify/respond…)渐进披露

灵枢(常驻校准器):
  · D_norm 验证闸门——记忆敢想,灵枢把关敢不敢说对
  · 诚实边界——不知道就说不知道(不裁决,防过度自信)
  · 自动补卡——agent 求助且无答案(双不会)→ 当场生成知识卡 → 下次即有

设计原则

  • 零 embedding:语义匹配为纯算法(词汇 n-gram / 共现桥 / 图传播),可解释、可审计;ONNX bge 仅作对比验证(显式配置启用)
  • 算法是参谋,agent 是主人:建边由 agent 决策,算法只给候选建议
  • 浅注入,深挖掘:每轮自动注入 3 行线索,细节由 agent 用工具渐进披露
  • 人类式学习:不会 → 求助 → 查不到 → 当场补卡(当日上限 5 张,防噪声);弱命中高频 → 自动补词网(俗语→规范词桥接)
  • 诚实边界:证据不足不裁决,图谱外明说,绝不硬答

功能清单

能力说明
记忆注入每轮 API 调度自动注入相关记忆入口(记忆+浅思维三行)
浅思维条件(邻域 kind 分布)/ 验证(灵枢校准)/ 边界(教训邻域+边界词)
自动补卡agent 求助无答案 → LLM 生成知识卡(空返回时启发式保底)→ 灵枢写入
自动补簇弱命中查询 ≥3 次 → 从查询提取俗语词自动建簇(零 LLM,相关性质量门)→ 词网自组织
占位卡治理无答案补卡的占位卡标 pending——不参与出招/验证(防污染),等 LLM 生成真卡
蒸馏保底回合后 LLM 蒸馏;空返回时启发式直接沉淀(记忆不断流)
自动成族语义图软簇(层次聚类,分辨率可调)——无硬分类
跨域桥agent 跨域对话触发桥边——「融会贯通」的结构化形态
灵枢自愈(v0.4.0)插件托管灵枢服务:未运行自动拉起 start_lingshu.py、崩溃按需重启(10s 冷却)、只 kill 自拉起进程(手动实例尊重)、不可用友好降级
过时记忆淘汰(v0.4.0)闲置 45 天 + 重要度门槛 → 自动退役(退出召回,search 可见 🗄,open/confirm 复活);蒸馏同题刷新/旧题接替;同标题只留最新;工具 engram_retire 手动处置
验证缓存(v0.4.0)灵枢验证 LRU(TTL 10min):同主题重复轮次零 HTTP;error 不缓存(恢复即重试)
工具面15 个工具:recall/store/propose/confirm/reject/open/search/link/update/remove/promote/status/verify/respond/retire

装配(与本地一致)

1. 克隆与构建

git clone https://github.com/yjh051108/dsh-engram-relay.git fengshi
cd fengshi
npm install --legacy-peer-deps   # 依赖精确锁定(package-lock.json 已入库)
npm run build            # 构建 lib/(tsc host + tsdown client)

2. 装配到 DSH

# 方式 A:热装(免重启,开发用)
# 用 DSH 注入器 dev_inject_plugin 指向本仓库目录

# 方式 B:正常装配(重启生效,生产用)
dsh plugin --profile web add .

3. 启动灵枢云(校准器服务,127.0.0.1:18766)

python lingshu/start_lingshu.py    # 自愈 watchdog:崩溃 1s 自动重启

v0.4.0 起可省略此步:插件(lingshuAutoStart=true)会在首次 verify/respond 时自动拉起服务;npm run verify 同样自动拉起(用完即停)。

4. 配置(cordis.patch.yml 或 schema 默认值)

config:
  lingshuVerifyUrl: 'http://127.0.0.1:18766'   # 灵枢校准器(默认开启)
  lingshuAutoStart: true                        # v0.4.0 融合自愈:未运行自动拉起
  lingshuPython: ''                             # 灵枢服务的 Python(空 = 沿用 pythonPath)
  retireEnabled: true                           # v0.4.0 过时记忆淘汰开关
  retireAfterDays: 45                           # 闲置超过此天数 → 自动退役(可复活)
  retireMaxImportance: 1                        # 仅重要度 ≤ 此值可自动退役(0 = 关)
  embedModel: ''                                # 空 = 纯算法语义匹配(默认)
  distillEveryTurns: 2                          # 回合蒸馏频率
  injectBudgetTokens: 200                       # 注入预算(token)

验证(装配后 30 秒确认)

npm run verify    # 运行自检:服务健康(未运行自动拉起)+ 卡库 + 记忆库 + 浅思维冒烟

新会话里问:

  • 「铁门放外面久了为什么生锈」→ 记忆/知识命中
  • 「量子纠缠能不能超光速通信」→ 浅思维验证行(?图谱外 或 ✓锚定)
  • 任意灵枢无卡的问题连续求助 → 自动补卡(~15s 生效)
  • 过时记忆治理:闲置记忆自动退役(engram_status 看 retiredCount)/ 手动处置 engram_retire / 误伤复活 engram_open 或 engram_confirm

与本地一致的说明

发布包包含:插件完整源码 + 灵枢运行集(lingshu/)+ 卡库种子 + 装配脚本(git clone 后需 npm run build 产出 lib/)。 本地与大家装配的差异仅剩:DSH 版本与 node_modules 解析(README 与 INSTALL 已覆盖)。

架构文档

  • docs/unified-brain.md:统一大脑架构(图网络+浅算子+校准器)
  • docs/unified-routing.md:零 embedding 统一路由设计(含文献:WordNet 扩展 / PRF 共现 / SPLADE)
  • docs/INSTALL-new-machine.md:换机装配指南

许可证

BSD-3-Clause(工程代码)。协议概念(智能论/信息差)权利归协议方。