Context Assembler

Not a tool for agents — a preprocessor for their context window.

OpenClaw injects ~3700 tokens of static bootstrap context into every session, regardless of the task. Context Assembler reduces that to ~400 tokens of task-relevant information, achieving 89% context compression without quality loss.

What It Does

Task arrives → Classify → Semantic Projection → Timeline Collapse → Forbidden Patterns → Pack → Agent reasons

1. Classifies the task type (nas_ops, coding, research, evolution_check, etc.)
2. Semantically projects relevant memory from MEMORY.md and daily notes — only what matters
3. Collapses timelines — repeated failures become single entries, noise gets filtered
4. Detects forbidden patterns — paths that failed ≥2 times get marked "do not retry"
5. Packs everything into a compact context block within token budget

Why This Matters

Karpathy's "Context Engineering > Prompt Engineering" principle applied to OpenClaw. The quality bottleneck isn't how you write prompts — it's what the agent sees in its context window before reasoning.

Quick Start

Manual invocation (agent calls it)

python3 scripts/assembler.py --task "check NAS disk health" --max-tokens 1500

Cron pre-turn hook

In your OpenClaw cron job, run assembler first:

# The agent's first action: get optimized context
python3 skills/context-assembler/scripts/assembler.py \
  --task "daily evolution check" \
  --max-tokens 1800

Agent-assisted mode

Tell your agent: "optimize my context" — it will call assembler and use the output.

Configuration

Edit genome.yml to customize:

• Source weights: which memory sources matter most
• Task profiles: per-task token budgets and source preferences
• Synonym map: lightweight semantic expansion (e.g., "shrimp" → "aquaculture, water quality")
• Noise patterns: timeline events to filter out
• Staleness decay: how fast old information loses relevance

The genome.yml is the "mutable kernel" — you tune it, the assembler engine stays fixed.

Requirements

• Python 3.8+
• PyYAML (pip install pyyaml)
• Read access to OpenClaw workspace (MEMORY.md, memory/*.md)

Architecture

skills/context-assembler/
├── SKILL.md              # This file
├── genome.yml            # ★ Mutable kernel (tune this)
├── scripts/
│   └── assembler.py      # Fixed engine (~510 lines)
├── index/                # Future: pre-built search indices
└── feedback/             # Selection → outcome log

Design Principles

1. Offload decisions: don't teach the agent to judge — encode judgment as a checklist
2. Compress output space: templates > free-form writing
3. Absence as signal: tell the agent what NOT to include
4. Embed domain knowledge: your expertise encoded as correlation rules
5. Graceful degradation: missing data is normal, not an error

Notes

• Phase 1 uses keyword matching with synonym expansion (zero-latency, zero-extra-memory)
• Phase 2+ will add embedding-based semantic search as memory corpus grows
• The genome is designed to be optimizable — feedback logging enables self-tuning
• Does NOT modify OpenClaw core — installs as a regular skill
• Contains no credentials, tokens, or personal identifiers — safe to publish