Memory Tools

Agent-controlled memory plugin for OpenClaw with confidence scoring, decay, and semantic search. The agent decides WHEN to store/retrieve memories — no auto-capture noise. v2 uses file-based storage with optional QMD search (no external APIs required).

Memory Tools v2

Agent-controlled persistent memory for OpenClaw.

What's New in v2

• File-based storage — Memories stored as readable markdown files
• No external APIs — No OpenAI dependency, everything runs locally
• QMD search (optional) — BM25 + vector + reranking via local GGUF models
• Auto-migration — Seamlessly upgrades from v1 SQLite/LanceDB storage

Why Memory-as-Tools?

Traditional memory systems auto-capture everything, flooding context with irrelevant information. Memory Tools follows the AgeMem approach: the agent decides when to store and retrieve memories.

Features

• 6 Memory Tools: memory_store, memory_update, memory_forget, memory_search, memory_summarize, memory_list
• Confidence Scoring: Track how certain you are (1.0 = explicit, 0.5 = inferred)
• Importance Scoring: Prioritize critical instructions over nice-to-know facts
• Decay/Expiration: Temporal memories automatically become stale
• Human-readable storage: Markdown files with YAML frontmatter
• Conflict Resolution: New info auto-supersedes old (no contradictions)

Installation

Step 1: Install from ClawHub

bash

clawhub install memory-tools

Step 2: Build the plugin

bash

cd skills/memory-tools
npm install
npm run build

Step 3: Activate the plugin

bash

openclaw plugins install --link .
openclaw plugins enable memory-tools

Step 4: Restart the gateway

bash

openclaw gateway restart

Optional: Install QMD for better search

bash

npm install -g @tobilu/qmd

Without QMD, basic filtering works. With QMD, you get semantic search (BM25 + vector + reranking).

Node Compatibility (QMD)

• QMD may not be compatible with the newest Node ABI immediately.
• If QMD fails (for example NODE_MODULE_VERSION mismatch), memory-tools falls back to basic mode.
• To force basic mode: MEMORY_TOOLS_DISABLE_QMD=true.
• For best QMD stability, use a Node LTS version supported by your installed QMD build.

Security Model

• No API keys or external credentials are required.
• Data is stored locally in ~/.openclaw/memories (or configured memoriesPath).
• The plugin can prepend standing-instruction context at before_agent_start when autoInjectInstructions=true.
• The plugin can auto-migrate legacy v1 data from ~/.openclaw/memory/tools/memory.db when autoMigrateLegacy=true.
• Defaults are conservative: both autoInjectInstructions and autoMigrateLegacy are false.

Storage Format

Memories are stored as markdown files in ~/.openclaw/memories/:

text

~/.openclaw/memories/
├── facts/
│   └── abc123-def4-5678-90ab-cdef12345678.md
├── preferences/
│   └── def456-7890-abcd-ef12-345678901234.md
├── instructions/
│   └── ghi789-0abc-def1-2345-678901234567.md
└── .deleted/
    └── old-memory.md

Each memory file:

markdown

---
id: abc123-def4-5678-90ab-cdef12345678
category: preference
confidence: 0.9
importance: 0.7
created_at: 2024-01-15T10:30:00Z
tags: [ui, settings]
---

User prefers dark mode in all applications.

Memory Categories

Category	Use For	Example
fact	Static information	"User's dog is named Rex"
preference	Likes/dislikes	"User prefers dark mode"
event	Temporal things	"Dentist Tuesday 3pm"
relationship	People connections	"Sarah is user's wife"
instruction	Standing orders	"Always respond in Spanish"
decision	Choices made	"We decided to use PostgreSQL"
context	Situational info	"User is job hunting"
entity	Named things	"Project Apollo is their startup"

Tool Reference

memory_store

text

memory_store({
  content: "User prefers bullet points",
  category: "preference",
  confidence: 0.9,
  importance: 0.7,
  tags: ["formatting", "communication"]
})

memory_search

text

memory_search({
  query: "formatting preferences",
  category: "preference",
  limit: 10
})

memory_update

text

memory_update({
  id: "abc123",
  content: "User now prefers numbered lists",
  confidence: 1.0
})

memory_forget

text

memory_forget({
  query: "bullet points",
  reason: "User corrected preference"
})

memory_summarize

text

memory_summarize({
  topic: "user's work projects",
  maxMemories: 20
})

memory_list

text

memory_list({
  category: "instruction",
  sortBy: "importance",
  limit: 20
})

CLI Commands

bash

# Show memory statistics
openclaw memory-tools stats

# List memories
openclaw memory-tools list
openclaw memory-tools list --category fact

# Search memories (requires QMD)
openclaw memory-tools search "dark mode"

# Export all memories as JSON
openclaw memory-tools export

# Force re-index with QMD
openclaw memory-tools reindex

# Show storage path
openclaw memory-tools path

Debugging

Memories are plain markdown files — just read them:

bash

cat ~/.openclaw/memories/facts/*.md
ls ~/.openclaw/memories/

Export all memories:

bash

openclaw memory-tools export > memories.json

Migration from v1

v2 automatically detects v1 databases and migrates them:

1. On startup, v2 checks for ~/.openclaw/memory/tools/memory.db
2. If found, exports all memories to markdown files
3. Original database preserved as backup
4. No manual action required

To enable auto-migration, set autoMigrateLegacy: true in plugin config.

License

MIT — Purple Horizons