Long-Term Memory

Overview

This skill gives OpenClaw agents long-term memory that persists across sessions. It replaces the manual MEMORY.md approach with automated capture + structured storage.

Everything runs locally — no cloud, no data leaves your machine.

Core Workflow

Session Start → 1. inject_context() → get relevant history
Session Run  → 2. remember() / auto_capture() → save important info
Session End  → 3. summarize() → compress session into memory

Scripts

scripts/memory_engine.py — Core engine

# Save a memory
python3 scripts/memory_engine.py remember "决定: 使用FastAPI框架" --tags decision,tech --importance 8 --project saas

# Search memories
python3 scripts/memory_engine.py search "技术方案" --tags tech --min-imp 5

# Get context for prompt injection
python3 scripts/memory_engine.py inject "当前任务描述..."

# Auto-capture from text (scans for decisions, facts, preferences)
python3 scripts/memory_engine.py auto "我们决定采用SQLite作为数据库，技术栈为FastAPI..."

# Session management
python3 scripts/memory_engine.py session-start    # returns session_id + context
python3 scripts/memory_engine.py session-end \x3Csession_id> --summary "..."

# Stats
python3 scripts/memory_engine.py stats

scripts/setup.py — One-time workspace setup

python3 scripts/setup.py

Memory Structure

• Storage: SQLite + FTS5 full-text search
• Fields: content, tags[], importance(1-10), source, session, project, timestamps
• Tags: Tag memories for filtering (e.g., decision, tech, user, project:X)
• Importance: 1-10 scale. 8+ = key fact, 6-7 = useful context, 1-5 = normal

Auto-Capture

The engine automatically detects important content from text:

AGENTS.md Integration

Add to your AGENTS.md (or the relevant agent's config):

## Long-Term Memory Rules

1. On session start: Run `python3 scripts/memory_engine.py inject "current task"` and use the output as context
2. When user shares important info: Use `remember()` to save it
3. Track decisions: Save key decisions with `--importance 8` and tag `decision`
4. Before answering "remember" or "previous" questions: Search memory first
5. On session end: Summarize key outcomes for next session

Data Storage

~/.openclaw/workspace/long-term-memory/
├── memory.db          # SQLite database
├── config.json        # Configuration
└── current_context.md # Last built context (for debugging)

Tips

• Be selective: Not everything needs remembering. Save decisions, preferences, problems.
• Use tags: project:X tags make cross-project memory searchable.
• Importance matters: 8+ for permanent facts, 5-7 for useful context, 3-4 for temporary.
• Search before answering: If user asks "do you remember X?", search memory first.