@smyslenny/agent-memory-openclaw

AgentMemory native adapter plugin for OpenClaw.

Automatically boots memories on session start and surfaces relevant memories into agent prompts — no manual MCP/cron orchestration needed.

Install

npm install @smyslenny/agent-memory-openclaw
# peer dependency
npm install @smyslenny/agent-memory

Configure

Add to your openclaw.json:

{
  "plugins": {
    "entries": {
      "@smyslenny/agent-memory-openclaw": {
        "enabled": true,
        "config": {
          "autoBoot": true,
          "autoSurface": true,
          "surfaceLimit": 5,
          "surfaceMaxChars": 2000
        }
      }
    }
  }
}

What It Does

Phase 1 (Current): Read Path

• Boot (session_start): Loads identity and pinned memories when a session starts. The narrative is cached and injected once via appendSystemContext.
• Surface (before_prompt_build): Retrieves relevant memories based on the user's message and recent conversation turns. Results are injected via prependContext.
• Dedup: Same memory is never injected twice within a session.
• Cleanup (before_reset): Clears session caches on /new or /reset.

Phase 2 (Current): Write Path + Feedback

• Remember (agent_end): Automatically extracts conversation content from successful agent turns and writes to long-term memory via rememberMemory().
  • Conservative thresholds: min 3 user turns, min 200 chars
  • 30-minute cooldown per session
  • Full provenance: source_session, source_context, observed_at
  • Write Guard handles dedup/merge/conflict
• Passive Feedback (before_prompt_build): Records recordPassiveFeedback() for surfaced memories, improving future retrieval ranking.

Config Options

Option	Type	Default	Description
enabled	boolean	true	Enable/disable the adapter
autoBoot	boolean	true	Auto-boot on session start
autoSurface	boolean	true	Auto-surface before each reply
autoRemember	boolean	false	Auto-remember after turns
autoFeedback	boolean	true	Record passive feedback for surfaced memories
surfaceLimit	number	5	Max memories to surface
surfaceMaxChars	number	2000	Max chars in injected prompt block
dbPath	string	auto	Explicit path to AgentMemory SQLite DB
rememberMinTurns	number	3	Min user turns to trigger remember
rememberMinChars	number	200	Min user chars to trigger remember
rememberCaptureRoles	string	"user-and-assistant"	Capture user-only or both
rememberCooldownMs	number	1800000	Cooldown between writes (30min)

DB Path Resolution

1. config.dbPath (explicit)
2. {workspaceDir}/agent-memory.db (from OpenClaw hook context)
3. AGENT_MEMORY_DB_PATH environment variable
4. ~/.agent-memory/memory.db (default)

Architecture

OpenClaw Gateway
  │
  ├── session_start ──→ BootManager ──→ boot()
  │
  ├── before_prompt_build ──→ SurfaceManager
  │                           ├── boot narrative (appendSystemContext)
  │                           ├── relevant memories (prependContext)
  │                           └── recordPassiveFeedback()
  │
  ├── agent_end ──→ RememberManager ──→ rememberMemory()
  │                 ├── conversation normalization
  │                 ├── threshold checks (turns, chars, cooldown)
  │                 └── provenance (session, context, timestamp)
  │
  └── before_reset ──→ clear all session caches
                         │
                         ▼
                   AgentMemory Core (v5.1+)

Requirements

• OpenClaw v2026.3.x+
• AgentMemory v5.1.0+
• Node.js 20+

License

MIT