HME — 给 DeepSeek 装上大脑

DeepSeek Harness 很强，但有个硬伤：会话结束，记忆清零 ，下次从头再来。

HME（Harness-Memory-Evolution）就是来治这个的。

鲸宝的三重记忆

1. 你是谁 —— 你的身份、偏好、习惯，跨项目记住你这个人。
2. 这个项目是什么 —— 每个项目不可替代的技术栈、约定、核心要点。
3. 你踩过的所有坑 —— 优秀经验、核心方法论，只要真有用，就往里记。

但它不臃肿。核心记忆用有界上限锁死、只留精华；海量经验沉进档案层，按需检索、满时提醒清理——就像 CPU 的多级缓存：小而快的常驻，大而全的按需取。

借鉴 Codex、Claude Code、OpenClaw、Hermes Agent 等成熟记忆系统的优点，让鲸宝不会失忆，也不会被杂事烦扰。

和别家比有什么不同？

维度	Codex CLI	Claude Code	OpenClaw	HME
记忆结构	MEMORY + USER 两文件	CLAUDE.md + skills	分层记忆	USER + MEMORY + archive 三层
防膨胀	硬上限（拒绝式）	无上限	向量库	有界 core + 扩容 archive
检索	无（全量注入）	skills 按需	向量检索	recall 关键词检索
经验进去	模型主动写	人类手写	自动	模型主动写 + 4 条蒸馏判据
可审计	纯文本	纯文本	结构化/黑盒	纯文本 § 分隔

30 秒装上

从 GitHub 装（推荐，立即可用）：

dsh plugin --profile web add github:weopenfire-git/hme-plugin

或从 npm 装（发布后可用）：

dsh plugin --profile web add @ymw/dsh-hme

然后补一行加载配置

打开 profile 的 `cordis.patch.yml`，加一条 host 行：

- insert:
    - id: hme-plugin
      name: '@ymw/dsh-hme'  # npm 装用包名；GitHub 装改为 github:weopenfire-git/hme-plugin
      config:
        memoryCharLimit: 2584
        userCharLimit: 1597

给你的 DeepSeek 装上大脑。

---

hme-plugin — Harness-Memory-Evolution

Give your DeepSeek a mind of its own.

DeepSeek Harness is formidable — yet it harbors one flaw: the moment a session ends, the memory resets. Preferences, conventions, hard-won lessons all evaporate, and the next session starts from blank.

Enter HME (Harness-Memory-Evolution).

The Whale's Threefold Memory

1. Who you are — your identity, your preferences, your habits, remembered across every project.
2. What this project is — the irreplaceable tech stack, conventions, and core points of each workspace.
3. The pits you've fallen into — proven lessons and core methodology; anything genuinely useful, commit to memory.

Yet it never grows bloated. Core memory is sealed by bounded caps to keep only the essence; the flood of experience sinks into an archive tier, retrieved on demand and pruned by suggestion when full — like a CPU's multi-level cache: the small and fast stay resident, the vast stay within reach.

Drawing on the strengths of Codex, Claude Code, OpenClaw, and Hermes Agent, HME lets your whale never forget — and never be distracted by trivia.

How it stacks up

Dimension	Codex CLI	Claude Code	OpenClaw	HME
Memory shape	MEMORY + USER two files	CLAUDE.md + skills	layered memory	USER + MEMORY + archive three-tier
Against bloat	hard cap (reject)	unbounded	vector store	bounded core + elastic archive
Retrieval	none (full injection)	skills on demand	vector search	recall keyword search
Expertise in	model writes manually	human hand-writes	automatic	model writes + 4 distillation criteria
Auditability	plain text	plain text	structured / black-box	plain text §-delimited

Install in 30 seconds

From GitHub (recommended, works right now):

dsh plugin --profile web add github:weopenfire-git/hme-plugin

Or from npm (once released):

dsh plugin --profile web add @ymw/dsh-hme

Then add one loader row

Open the profile's cordis.patch.yml and add a host row:

- insert:
    - id: hme-plugin
      name: '@ymw/dsh-hme'  # package name for npm; use github:weopenfire-git/hme-plugin for GitHub
      config:
        memoryCharLimit: 2584
        userCharLimit: 1597

Give your DeepSeek a mind of its own.

---

Technical reference

Cross-session long-term memory for DeepSeek Harness, as an out-of-tree plugin. It gives an agent two bounded plain-text memory files plus a memory tool, so user preferences and project facts survive across sessions instead of being relearned every time.

Status: Phase 1 + Phase 1.5 — core memory plus the archive overflow layer. Developer preview: the Harness APIs this plugin builds on are still changing. See DESIGN.md and ARCHIVE.md.

What it ships

• A memory tool that adds, replaces, and removes durable facts.
• An archive tool (add/replace/remove/move) plus a recall tool over a larger per-workspace overflow layer.
• Two context blocks, frozen at session start and injected into every turn of that session:
  • hme:user — the global user profile.
  • hme:memory — the current workspace's project facts.

Memory files

File	Location	Hard cap
USER.md	$DSH_HOME/hme/USER.md (global, shared across workspaces)	1,597 chars (F₁₇)
MEMORY.md	<workspace>/.dsh/hme/MEMORY.md (per workspace)	2,584 chars (F₁₈)

Facts are one per line, §-terminated, and human-editable. Caps count Unicode code points. A write that would exceed the cap is rejected, never truncated. Every write re-reads the latest file and merges at the fact level, so concurrent sessions do not clobber each other's edits.

Install

Install from npm or Git into a profile, then reference it by package name:

dsh plugin --profile web add @yinging/dsh-hme
# or from Git: dsh plugin --profile web add github:weopenfire-git/hme-plugin

Add a host row to the profile's cordis.patch.yml (web, CLI, and headless alike):

- id: hme-plugin
  name: '@yinging/dsh-hme'
  config:
    memoryCharLimit: 2584
    userCharLimit: 1597

userMemoryFile and workspaceMemoryFile have sensible defaults; override them to relocate the files.

The memory tool

Parameter	Type	Required	Meaning
action	add | replace | remove	yes	Mutation to apply.
target	memory | user	yes	memory → workspace MEMORY.md; user → global USER.md.
content	string	for add / replace	New fact text.
old_text	string	for replace / remove	Substring that locates the fact to change or drop.

replace and remove locate a fact by substring; if several facts match, the tool reports ambiguity and asks for a narrower old_text. A write never changes the current session's injection — the next session re-reads the files.

Configuration

Field	Default	Meaning
memoryCharLimit	2584	Hard character cap for MEMORY.md (F₁₈).
userCharLimit	1597	Hard character cap for USER.md (F₁₇).
userMemoryFile	<dshHome>/hme/USER.md	Global user-profile path.
workspaceMemoryFile	.dsh/hme/MEMORY.md	Workspace-relative project-memory path.
archiveCharLimit	131072	Soft character cap for the archive (2¹⁷ ≈ one context window).
archiveMemoryFile	.dsh/hme/archive.md	Workspace-relative archive path.

How it works

At agent/session-start the plugin freezes both files into a per-scope snapshot, and the two context blocks serve that snapshot for the whole session. Writes go straight to disk with read-merge-write, so the current session's view stays stable while the next session sees the latest facts. Injection uses systemPrompt.context(), which records the memory as a persistent user-role snapshot rather than re-reading the files each turn.

Develop / verify

hme-plugin is a pnpm workspace member for dependency resolution only, not a build target. Run pnpm install once so its @deepseek-ai/* dependencies link, then:

pnpm exec tsc -p hme-plugin
pnpm exec vitest run --config hme-plugin/vitest.config.ts

Roadmap

• Phase 1.5 — archive overflow layer (archive / recall tools); see ARCHIVE.md.
• Phase 2 — skill evolution: generate, optimize, and curate skills.
• Phase 3 — reflection and maintenance in background jobs.

License

MIT