Show and tell: dsh-agent-memory — self-evolving memory with a "dream" pass, user profile, and provenance from the replayable session log（自进化记忆插件）

[findshan]

TL;DR

@dsh-kit/agent-memory — a self-evolving memory plugin for DeepSeek Harness. It brings the industry's "dream" memory-consolidation pattern (Claude Code Dream / ChatGPT Dreaming / Letta sleep-time) to DSH for the first time, and it builds on the one asset Claude Code and Codex structurally lack: the replayable, typed session log.

Repo: https://github.com/findshan/dsh-agent-memory · npm: @dsh-kit/agent-memory · MIT · dsh-plugin topic

Why memory, why now

We surveyed the DSH memory ecosystem (120+ plugins) and the industry (Claude Code, Codex, ChatGPT Dreaming, Letta, MEM0, Zep) and found two things:

1. The "dream" is now an industry-standard mechanism — background memory consolidation while the user is away (merge, dedupe, correct, prune). Anthropic's /dream is shipped, OpenAI's Dreaming V3 is live, Letta has sleep-time agents. But zero DSH plugins implement it.
2. DSH has a structural advantage nobody is using: every session is a fully typed, replayable event log — user/message with source, tool/call/tool/result with exact args and results, assistant/message with token usage, request/header snapshots, fork/resume lineage. Claude Code's memory system has to grep JSONL transcripts; DSH memory can be derived from ground truth with provenance back to exact (sessionId, seq) events.

What it does

Capture → Dream (consolidate) → Retrieve/inject → Evolve

• Layered belief store: user (profile) / project / session / global scopes, each a plaintext belief with confidence, importance, and provenance.
• Correction learning: you correct the agent → it learns automatically → the new belief supersedes the old one (history kept, auditable). The #1 trust builder, built into the write path.
• The Dream pass: a gated background job (interval + session count + lock file) that reads new session events, merges duplicates, supersedes contradictions, archives stale beliefs, and refreshes a "last time we were doing X" resume digest — injected at session start so the agent opens by continuing, not by asking.
• Frozen profile snapshot: top user + project beliefs injected at ≤2k tokens (PersonaMem-v2 shows 2k-token agentic memory beats full-context personalization while using 16× fewer tokens).
• 7 compact tools for the model: memory_save / memory_search / memory_list / memory_confirm / memory_forget / memory_profile / memory_dream.
• Governance: writes land as suggested and only become active on human confirmation; everything is plaintext, local-first, zero native dependencies.

The 30-second demo

dsh plugin --profile web add @dsh-kit/agent-memory

// any plugin: inject: ['memory']
await ctx.memory.remember({ content: '用户偏好 pnpm 而非 npm', scope: 'user', kind: 'preference' })
const hits = await ctx.memory.search('pnpm')   // BM25, ranked by relevance × confidence × importance
await ctx.memory.confirm(id)                   // human confirms → active
await ctx.memory.dream(true)                   // force a consolidation pass

Where the research went

We wrote the full landscape survey while designing this — competitor mechanisms, memory architectures (MEM0/Letta/Zep), the DSH ecosystem gap analysis, and the mechanism verification against DSH source. It's in the repo: research/ + PRD.md.

Why this position is open

Position	Who holds it today
Three-in-one (self-evolving + profile + project)	dsh-memory-evolve — but fully self-built storage, no official infra
Official-infra reuse (approval, systemPrompt, sessionQuery)	dsh-memento / Max-Null — but no evolution/dream
Log-citation extraction	Jesse-njx/dsh-memory — but no profile, no dream
All of it + dream consolidation + provenance from the replayable log	nobody — this is the open seam we're claiming

Feedback wanted: is the 7-tool surface right? Should the dream's LLM-assisted consolidation (v0.2) use a forked subagent like AutoDream, or inline cheap-model calls? What would make this the memory seam you'd actually adopt?

---

中文版

@dsh-kit/agent-memory —— 为 DeepSeek Harness 打造的自进化记忆插件。首次把行业前沿的 「梦境式记忆整合」（Claude Code Dream / ChatGPT Dreaming / Letta sleep-time）带到 DSH，并且建立在 Claude Code 和 Codex 结构上做不到的资产上：可回放、类型化的会话日志。

为什么现在做：调研了 DSH 生态 120+ 记忆插件与全行业后确认——①「梦境」已是行业标准机制，但 DSH 生态零实现；② DSH 的全量类型化事件日志（tool/call 精确参数、assistant/message 带 token 计费、fork 谱系）是别人没有的 ground truth，却没人用它做记忆。

它做什么：捕获 → 梦境整合 → 检索注入 → 进化。

• 分层信念库（用户画像 / 项目 / 会话 / 全局），每条带置信度、重要性、溯源
• 纠错即学：纠正 → 自动学习 → 覆盖旧信念（历史可审计）
• 梦境 pass：门控后台任务，合并去重、淘汰过时、刷新"上次进展"叙事
• 冻结画像注入 ≤2k token；7 个精简工具；suggested→active 人机共治；明文本地、零原生依赖

仓库：https://github.com/findshan/dsh-agent-memory · 安装：dsh plugin --profile web add @dsh-kit/agent-memory

期待大家的反馈：工具面是否合适？v0.2 的 LLM 整合该用 fork 子代理（AutoDream 式）还是内联轻量模型？怎样才是你愿意真正采用的记忆缝？

[findshan]

Update: real-model test with deepseek-v4-flash ✅

Ran the full loop against the DeepSeek API with the real plugin and the real 7 tools (no stubs). Transcript: test/integration.mjs in the repo. Highlights:

1. The model uses the tools on its own — user states a preference, the model calls memory_save with correct scope/kind classification:

⚙️ memory_save({"content":"用户开发项目时使用 pnpm 作为包管理器，而不是 npm。","scope":"user","kind":"preference"})

2. Cross-session recall works — brand-new session, zero context, durable memory answers:

用户: 我上次说过我用什么包管理器吗？→ memory_search hit → 助手: 是的，你之前提到过：你开发项目时使用 pnpm 作为包管理器，而不是 npm。

3. Correction learning prevents repeat mistakes — one correction, new session, model applies the new preference and the old belief is archived (not deleted, auditable):

纠正 → old pnpm belief superseded → 助手: 根据记录，你现在用的是 npm（之前改用 npm 后就不再使用 pnpm 了）。

4. Governance holds — write lands suggested → human confirm → active; stats show { total: 2, active: 1, archived: 1, correctionCount: 1, medianConfidence: 0.95 }.

Full write-up (Chinese): BLOG.md

Next: v0.2 LLM-assisted dream consolidation (forked read-only subagent, AutoDream-style four phases), adoption-feedback loop, Ebbinghaus decay. Feedback welcome!