Self Improvement Loop

self-improvement-loop v4.6.13

Overview

Core Capability: Each agent has its own self-improvement feedback loop — isolated learnings directory, independent cron scan, notifications bound to their own channel bot, and A/B/C/D execution in the correct agent session.

User corrects code-dev agent
    ↓
code-dev hook captures → writes to code-dev/.learnings/
    ↓
self-improvement-code-dev cron (hourly) → scans only code-dev/.learnings/
    ↓
Pattern detected (count≥2) → notifies user via code bot
    ↓
User replies A → code-dev agent session executes → writes back to code-dev workspace

Data Flow (Per-Agent Isolation)

openclaw.json (agents.list + bindings)
    │
    ├─→ install.sh reads agents
    │         ↓
    │    Creates per-agent directories (e.g., agents/workspace/.learnings/)
    │
    ├─→ install.sh reads channel accounts
    │         ↓
    │    Writes bindings (agentId → accountId)
    │    { "agentId": "code-dev", "match": { "channel": "telegram", "accountId": "code" } }
    │
    └─→ setup_crons.py creates per-agent cron
              ↓
         Each cron carries --agent flag
         LEARNINGS_DIR points to that agent's workspace
         delivery uses that agent's account

┌─────────────────────────────────────────────────────────────┐
│  Hook (handler.js) - Global shared, dynamic routing         │
│  sessionKey: "agent:code-dev:telegram:..."                 │
│  → extractAgentId("code-dev")                              │
│  → getAgentWorkspace("code-dev") → agent workspace path     │
│  → learningsDir = workspace + "/.learnings/"               │
│  Pushes reminders to context.messages                      │
└─────────────────────────────────────────────────────────────┘

┌─────────────────────────────────────────────────────────────┐
│  Cron (per-agent) - Independent execution                   │
│  Name: self-improvement-code-dev                           │
│  --agent code-dev → bound to code-dev agent session        │
│  LEARNINGS_DIR=".../agents/code-dev/.learnings"             │
│  delivery: channel=telegram, accountId=code                │
│  Scans only its own LEARNINGS_DIR                          │
└─────────────────────────────────────────────────────────────┘

┌─────────────────────────────────────────────────────────────┐
│  Notification - Sent via respective channel bot           │
│  code bot → user conversation with code-dev                 │
│  moling bot → user conversation with moling                 │
│  User replies A/B/C/D → OpenClaw routes to correct agent   │
└─────────────────────────────────────────────────────────────┘

┌─────────────────────────────────────────────────────────────┐
│  A/B/C/D Execution - Runs in correct agent session        │
│  Reads pending JSON → executes skill-creator/skill-improvement
│  Writes back to that agent's learnings directory            │
└─────────────────────────────────────────────────────────────┘

Permissions Required

Security Boundaries:

• File operations strictly scoped to per-agent learnings directories under ~/.openclaw/workspace/
• All paths dynamically generated, no hardcoding
• Cron runs in isolated session with minimal toolset
• install.sh prompts for confirmation before writing bindings, can be skipped

Core Components

Installation

bash ~/.openclaw/workspace/skills/self-improvement-loop/install.sh
openclaw gateway restart

Installation automatically:

1. Creates .learnings/ directories for each agent
2. Installs hook and scripts
3. Creates per-agent cron jobs
4. Updates openclaw.json bindings (enables message routing)
5. Injects self-improvement index instructions into each agent workspace's AGENTS.md and memory.md

A/B/C/D handling logic is stored in skills/self-improvement-loop/scripts/agents-append.md (shared). During install, index references are injected into each agent's AGENTS.md and memory.md:

## Self-Improvement (A/B/C/D)

Before handling A/B/C/D, must consult:
skills/self-improvement-loop/scripts/agents-append.md

No manual appending needed after install; skill updates automatically use the latest version for all agents.

Verification

# Check bindings
python3 -c "import json; print(json.dumps(json.load(open('$HOME/.openclaw/openclaw.json')).get('bindings',[]), indent=2))"

# Check per-agent crons
openclaw cron list | grep self-improvement

# Test distill (specific agent)
LEARNINGS_DIR="$HOME/.openclaw/workspace/agents/code-dev/.learnings" \
  bash ~/.openclaw/workspace/skills/self-improvement-loop/scripts/distill.sh --check-only

Dependencies

Limitations & Known Issues

• ACP runtime agent (claude): No independent channel, falls back to defaultAccount
• API (Gateway): API unreachable (404), falls back to CLI for cron creation, sessionTarget downgraded to isolated
• Concurrent notifications: When multiple agents trigger simultaneously, user replies may route to wrong agent session
• sessionTarget=current: Requires Gateway API support, currently unavailable

See Also

• references/setup-guide.md — Full installation and configuration guide