Plugin directory / AI Models / dsh-agent-memory
dsh-agent-memory
Unverified findshan
What it does
Self-evolving memory for DeepSeek Harness: capture, dream consolidation, retrieval injection, evolve. Includes user profile, project memory, correction learning, resume narrative with provenance.
Unverified — not yet verified
Self-evolving memory for DeepSeek Harness: capture, dream consolidation, retrieval injection, evolve. Includes user profile, project memory, correction learning, resume narrative with provenance. Not yet verified — install and test it yourself.
“Unverified” means our automated CI has not yet installed this plugin. Feature descriptions and version compatibility are the author’s claims. This is not a security audit and not an endorsement of third-party code.
README
dsh-evolving-memory
文件化自进化记忆 —— DeepSeek Harness 的跨会话压缩层(v2)
记忆的本质是信息的压缩与提取。本插件叠在官方 compaction(会话内压缩)之上,用廉价模型把已压缩的会话摘要进一步提取、整合为人类可读的 Markdown 记忆文件——用户可看、可改,智能体可检索、可进化。
设计要点
| 记忆 = 压缩与提取 | 消费官方 compaction/summary 事件做提取(T1),不重复读原始日志;dream 做跨会话整合(T3) |
| 纯 Markdown 文件 | user / agent / memory / dream + 每项目 project.md + daily/ 时间层;无 schema、无迁移、无置信度、无状态机 |
| 人机共治 | user.md 用户可读可编辑;agent 写入只是建议,用户有最终决定权 |
| 目录式披露注入 | 常驻注入只有记忆目录(确定性生成,零模型调用,有界);内容按需 memory_search(找)+ memory_read(展开)披露——与 DSH skill 机制同构 |
| 三层压缩 | 会话 compact(官方)→ 提取进 daily/(情景记忆)→ dream 固化进主题文件(语义记忆),对应"睡眠时情景→语义固化" |
| 廉价模型驱动 | 提取与整合走便宜模型(KB 级输入一次调用);无 key 时记忆读写/检索照常,仅模型步骤跳过 |
Install / 安装
dsh plugin --profile web add dsh-evolving-memory
dsh --profile web
记忆文件落在 $DSH_HOME/memory/:
├── daily/2026-08-16.md # 今日要点 + 会话纪要 + 待办(时间层,dream 后归档)
├── user.md # 用户画像(身份/偏好/目标/禁忌/想法)—— 用户可编辑
├── agent.md # 智能体自我认知与工作方式
├── memory.md # 长期记忆(兜底落点)
├── dream.md # 梦境整合日志(审计)
└── projects/<p>/project.md
工具(6 个)
memory_search(找)· memory_read(展开)· memory_catalog(目录)· memory_save(写)· memory_correct(纠错即学)· memory_dream(整合)
分类是软约定(写进 prompt,不是代码):用户的事→user.md,项目的事→project.md,时间→daily/,拿不准→memory.md。检索兜底一切——放错文件也能找到。
配置
| Key | 默认 | 含义 |
|---|---|---|
memoryDir |
$DSH_HOME/memory |
记忆根目录 |
dreamIntervalHours / dreamMinSessions |
24 / 5 |
梦境整合门控 |
model / apiKey / baseURL |
deepseek-chat / env / api.deepseek.com |
廉价模型(提取+整合) |
catalogBudgetTokens / catalogTopN |
1000 / 5 |
目录注入预算与每文件行数 |
searchTopK |
5 |
默认检索条数 |
autoExtract |
true |
消费 compact summary 自动提取 |
dailyRetentionDays |
30 |
daily 保留天数,dream 后归档 |
Development / 开发
npm run typecheck
npm run build
node test/smoke.mjs # 无模型:布局/保存/读取/检索/目录/纠错/捕获/时间层/持久化
DEEPSEEK_API_KEY=sk-... node test/integration.mjs # 真实模型:提取/纠错/梦境整合
为什么做这个
DSH 生态的记忆插件要么是 JSON 记录 + 置信度机器(复杂),要么没有跨会话整合。v2 回归本质:记忆就是几份活的文档 + 一个会读写的智能体——系统只提供文件、检索器、整合触发器,判断全部交给模型。明文本是 AGI/ASI 时代的通用接口。PRD:PRD.md · 旧方案:PRD-v1.md
License
MIT
Install
Install the catalog once, then DeepSeek Harness can find and install any plugin from this site automatically:
dsh plugin add dshbase-catalog Then say "install dsh-agent-memory for me" — your agent finds it in the directory and installs it. Docs: dshbase-catalog · verified packs.
This plugin is GitHub source (not published to npm) — install it straight from the repo:
Web profile:
dsh plugin --profile web add github:findshan/dsh-agent-memory Headless (CLI) profile:
dsh plugin --profile headless add github:findshan/dsh-agent-memory Test report
Not yet L3-verified — see failure note below if we already ran it.
Note: 验证: runtime-fail (0.1.0-rc.6) Browse all pending failures →
When to use it
Bring a new model, provider, or routing policy into the loop so dsh can pick the right brain for the job.
Who it's for
Users juggling multiple models or providers who want cost, quality, and latency balanced automatically.
For developers — extending it
Provider adapters and routing heuristics are the seams — add a backend, tune the fallback chain, or add per-task model selection.