lossless-claw — Lossless Context Management

What

lossless-claw is an OpenClaw plugin that replaces the built-in sliding-window compaction with a DAG-based summarization system. Instead of truncating older messages when context fills up, it:

  1. Persists every message in a SQLite database
  2. Summarizes chunks of older messages into compressed nodes
  3. Condenses summaries into higher-level nodes as they accumulate (forming a DAG)
  4. Assembles context each turn by combining summaries + recent raw messages
  5. Provides agent tools (lcm_grep, lcm_describe, lcm_expand) so the agent can search and recall details from compacted history

Why we install it

The problem: Without LCM, OpenClaw’s default compaction simply drops older messages when context fills up. In long conversations (multi-hour sessions, daily check-ins), context gets lost — the agent forgets decisions made hours or days ago.

The solution: LCM ensures nothing is ever lost. The agent can drill into any summary to recover original details. For a personal assistant that needs to remember ongoing projects, house hunting progress, business decisions — this is critical.

In practice: The difference between “my AI forgot what we discussed yesterday” and “my AI remembers everything.” Essential for long sessions, daily check-ins, or any workflow where continuity across conversations matters.

Cost: Minimal. Summarization uses Anthropic Haiku (~$0.001-0.003 per compaction pass). For typical usage, adds ~€1-5/month. Configured via LCM_SUMMARY_MODEL to use a cheap model, not the main conversation model.

Configuration

openclaw.json

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17

{
  "plugins": {
    "entries": {
      "lossless-claw": {
        "enabled": true,
        "config": {
          "freshTailCount": 32,
          "contextThreshold": 0.75,
          "incrementalMaxDepth": -1
        }
      }
    },
    "slots": {
      "contextEngine": "lossless-claw"
    }
  }
}

Key settings explained

SettingValueWhy
freshTailCount32Protects last 32 messages from compaction — enough recent context for continuity
contextThreshold0.75Triggers compaction at 75% of context window — leaves headroom for responses
incrementalMaxDepth-1Unlimited DAG depth — summaries cascade as deep as needed automatically
LCM_SUMMARY_MODEL (env)anthropic/claude-haiku-4-5Direct Anthropic API (not OpenRouter) — cheapest option for summarization. Set in .env.

Installation

The plugin is installed automatically via the Docker entrypoint script on every container start. It is enabled by default — no action needed to opt in.

To disable it, set INSTALL_LOSSLESS_CLAW=0 in secrets/.env:

1

INSTALL_LOSSLESS_CLAW=0

Then redeploy: make deploy REBUILD=1.

Installation is idempotent — safe to run repeatedly. No Dockerfile changes needed.

References