Context Optimizer

Keep AI agents from forgetting. Keep context small.

What It Does

A dual-memory architecture that keeps your LLM context under 500 tokens while preserving critical information:

• Hot Layer (<500 tokens): Recent summaries, decisions, current tasks
• Cold Layer (unlimited): Full history, vector-indexed for retrieval

NEW in v2.0: Automatic triggering via message:received hook!

Why

LLM applications suffer from:

• Context window overflow
• Memory loss between sessions
• Token explosion with long conversations
• No synchronization across agents

Context Optimizer solves these with automatic summarization and smart memory分层.

Quick Start

# Install
clawhub install agent-context-optimizer

# That's it! Hook auto-installs and enables
# Every user message triggers automatic summarize

Architecture

┌─────────────────────────────────────┐
│  HOT Layer (<500 tokens)            │
│  - Recent summaries                 │
│  - Decisions                       │
│  - Current tasks                   │
│  - Key facts                       │
├─────────────────────────────────────┤
│  COLD Layer (unlimited)            │
│  - Full conversation history        │
│  - Vector indexed for retrieval     │
│  - Loaded on-demand                │
└─────────────────────────────────────┘

How Auto-Trigger Works

User sends message
    ↓
Hook listens to message:received
    ↓
Checks: "5+ minutes since last summary?"
    ↓
Yes → Read transcript → Summarize → Update hot layer
    ↓
Agent always has fresh context

Files

Path	Purpose
hook/HOOK.md	Hook metadata
hook/handler.ts	Auto-summarize on every message
scripts/context-optimizer.js	Core engine
scripts/agent-memory-helper.js	Startup loader

Testing

# Check hook
openclaw hooks list | grep context

# Verify hot layer updates
cat ~/.openclaw/workspace/knowledge/agents/main/sidecar.md

License

MIT