imemory

中文文档:README.md

⚠️ Release status: Built on Aug 14, 2026 and put into daily use the very next day, this project has already gone through 4 iterations. Because it is highly personalized, the published version is still a beta — APIs and behavior may change as it evolves. Thanks for your understanding.

imemory is a self-evolving long-term memory plugin for DeepSeek Harness (DSH). It keeps your long-term memory in two plain-text files (MEMORY.md — your notes, and USER.md — the user profile; entries separated by a lone § line). After each turn, an LLM extracts durable facts worth remembering and writes them back; before each step, relevant memories are recalled and injected into the context. Reusable workflows are distilled into learned-* skills. A web panel is included.

Features

• Auto-extraction after each turn — the LLM mines the conversation for durable facts (importance 1–10; <3 is dropped; better few than many)
• Per-step recall injection — keyword pre-filter + LLM semantic re-rank (temperature 0), injected without announcing the memory use
• Session opener — new sessions start with the user profile + high-importance notes
• Conflict merge / dedup / decay — consolidated every 6 h (auto-merges plugin-owned entries; user-edited entries only get suggestions)
• Skill learning loop — reusable procedures become learned-* skills (skills.json → DSH skill registry, loadable via the skill tool)
• 6 tools — memory_search / memory_add / memory_update / memory_forget / memory_sessions / memory_consolidate
• Web panel — /imemory/ (browse entries / hit counts / importance up-down / delete / consolidate)
• Zero dependency — imports no @deepseek-ai package; loads from any install path (manual copy, link:, npm/git)
• Profile-agnostic — webServer is optional: web profiles get the panel automatically; headless/tui profiles run fine without it

Quick start (60 s)

# 1. copy the plugin into your profile dir (web profile example)
cp imemory.mjs ~/.dsh/profiles/web/imemory.mjs

# 2. append to ~/.dsh/profiles/web/cordis.patch.yml:
#    - insert:
#        - id: imemory
#          name: ./imemory.mjs

# 3. restart dsh. You should see: [imemory] 就绪 ... 记忆目录 ...

Installation

Target DSH: 0.1.0-rc.6 (Node ≥ 20). Two ways:

Method A — Manual mount (recommended, no build/npm)

1. Copy imemory.mjs into your profile dir, e.g. ~/.dsh/profiles/web/imemory.mjs
2. Append to the profile's cordis.patch.yml:

- insert:
    - id: imemory
      name: ./imemory.mjs
      # optional: config:
      #   memDir: ~/.dsh/my-memories

1. Restart dsh.

Method B — Bundle install (npm / git)

dsh plugin --profile web add imemory                    # after npm publish
dsh plugin --profile web add https://github.com/bass1125/imemory.git

The package's dsh.bundle.patch declaration mounts the plugin automatically. Remove any manual mount line from Method A first to avoid double-mounting.

Hot reload

• Config changes (e.g. memDir): the profile's cordis.patch.yml is watched (watchUserPatches), saving takes effect immediately.
• Plugin code changes: restart dsh; or force a re-import by bumping the module URL query in the insert line, e.g. name: ./imemory.mjs?v=2.

Configuration

Only one option: memDir (memory/notes directory).

The directory is created automatically on startup.

Memory layout

The harness home is $DSH_HOME, defaulting to ~/.dsh (Windows: C:\Users\<you>\.dsh). No hardcoded paths — works out of the box on any machine:

The memory files are plain text — edit them anytime. § on its own line separates entries; don't use lone § lines inside an entry.

Verification

• A new session's first step carries an <imemory> opener block
• 6 memory_* tools appear in the model's tool list
• Panel: http://<dsh-host>:<dsh-port>/imemory/ (web profiles)

Behavior notes

• Extraction: root sessions only; 20 s cooldown; ≤6 memory ops / 2 skills per round
• Recall: keyword scoring, ≤15 candidates go through LLM re-rank; plain keyword fallback without a model
• Consolidation: every 6 h — conflict merge + dedup + decay (entries >14 days with effective score ≤1.5 are retired) + index orphan cleanup
• Importance: 1–10; <3 dropped; hits raise the effective score
• Hermes-compatible: same §-separated format as Hermes memory files — point memDir at an existing Hermes memory dir to adopt old memories seamlessly

Development

imemory/
├─ imemory.mjs       plugin body (single file, zero build, zero deps)
├─ package.json      bundle declaration (dsh.bundle.patch)
├─ cordis.patch.yml  bundle mount patch
├─ README.md / README_EN.md
└─ LICENSE

License

MIT