dsh-memory

Long-term memory plugin for DeepSeek Harness (dsh). Automatically recalls relevant memories before answering, saves key facts after answering, and provides a visual settings page for full memory management.

Features

• Persistent memory store: JSON file under $DSH_HOME/memory-plugin/memories.json
• Auto-recall: System-prompt instructions tell the model to call memory_recall at the start of every answer
• Auto-save: Model is instructed to call memory_save when it learns worth-remembering facts
• Dynamic context: Recent high-importance memories (≥3 stars) are automatically injected into every request's runtime context
• Full CRUD: 5 model tools — save, recall (search), list, update, delete
• Smart search: Keyword + tag matching with importance and recency scoring
• Visual settings page: Browse, search, add, edit, delete memories in a dedicated settings section
• Import / Export: Backup and restore memories as JSON files (merge or replace mode)

Installation

From npm (recommended)

dsh plugin --profile web add dsh-memory

Then restart dsh (or the web UI).

From GitHub

dsh plugin --profile web add github:doublehappy123/dsh-memory

Usage

Model tools

The plugin registers 5 tools that the model can call automatically:

Tool	Description
memory_save	Save a fact/preference to long-term memory
memory_recall	Search memory by keywords/tags, returns ranked results
memory_list	List all memories (paginated, tag-filterable)
memory_update	Update an existing memory by ID
memory_delete	Delete a memory by ID

The system prompt instructs the model to:

1. Call memory_recall before answering to retrieve relevant context
2. Call memory_save after answering when new worth-remembering facts emerge

Visual settings page

After installation, open Settings → 长期记忆 (Long-term Memory) in the dsh web UI. You can:

• Browse all memories as cards with tags, importance stars, and timestamps
• Real-time search by content or tags (300ms debounce)
• Add new memories with content, tags, category, and importance (1-5 stars)
• Edit existing memories
• Delete memories (with confirmation)
• Paginate through large memory collections (20 per page)
• Export all memories to a downloadable JSON file
• Import memories from a JSON file (merge mode: dedupe by ID, update existing, add new)

Memory record

{
  "id": "uuid",
  "content": "self-contained fact or preference",
  "tags": ["preference", "python"],
  "category": "technical",
  "importance": 4,
  "createdAt": "2026-09-01T12:00:00.000Z",
  "updatedAt": "2026-09-01T12:00:00.000Z",
  "source": "model"
}

Importance levels

Level	Meaning
1	Ephemeral — likely irrelevant soon
2	Low — minor preference or context
3	Normal — default, useful recurring info
4	High — important preference or decision
5	Critical — must never be forgotten

Memories with importance ≥3 are automatically injected into the model's context at the start of every request.

Data location

• Memory store: ~/.dsh/memory-plugin/memories.json
• Plugin log: ~/.dsh/logs/dsh-memory/dsh-memory.log

Import / Export format

Exported files have the following structure:

{
  "version": 1,
  "exportedAt": "2026-09-02T...",
  "count": 42,
  "memories": [ ... ]
}

Import accepts either this format or a raw array of memory objects.

Requirements

• dsh >= 0.1.1-rc.2
• Node.js >= 20
• Web profile (for the visual settings page; tools work on headless profiles too)

License

MIT