Claw Compactor — workspace token compression

What it does

Claw Compactor provides a multi-layer pipeline to compress workspace content and reduce LLM token costs. It combines deterministic rule-based compression (dedup, dictionary encoding, RLE, tokenizer optimizations) with an optional Engram LLM-driven Observational Memory layer for semantic summarization and long-term reflections. Includes CLIs to run benchmarks, full pipelines, and targeted layer runs.

When to use it

Run at session start to shrink injected system context, before expensive LLM calls to lower cost, or as a regular maintenance cron to compress accumulated memory files. Use benchmark mode to estimate savings; run full pipeline when savings justify writes. Useful for teams managing large agent workspace histories.

What's included

• Scripts: mem_compress.py (entry point), engram_cli.py, engram_auto.py and various helpers.
• References: configuration examples (engram.yaml), thresholds guidance, and output artifact layout.
• Instructions: CLI usage for full pipeline, per-layer commands, Engram API examples, daemon mode and integration tips for injecting context into system prompts.

Compatible agents

Agents with workspace access and ability to run Python/CLI (OpenClaw, CLI-capable automation agents). Engram layer requires an LLM-compatible endpoint or API key for Anthropic/OpenAI-compatible services.