- Blog
- Say Goodbye to Openclaw Amnesia: A Deep Dive and Complete Guide to lossless-claw
Say Goodbye to Openclaw Amnesia: A Deep Dive and Complete Guide to lossless-claw
Table of contents
Say Goodbye to Openclaw Amnesia: A Deep Dive and Complete Guide to lossless-claw
Welcome back to our tech sharing session. If you frequently work with AI agents, you are likely familiar with the dreaded context window limit. When a conversation gets too long, traditional systems are forced to either delete old messages or compress them into a rough summary. That is how agents lose crucial details, active task state, and important nuance.
Today we are looking at a very different approach: lossless-claw. This guide covers the technical principles behind lossless-claw, the practical advantages it brings to AI engineering, and the exact steps required to get lossless-claw running in OpenClaw.
The Core Principles of lossless-claw
Instead of waiting for the context window to overflow and then scrambling to recover, lossless-claw works continuously and asynchronously in the background.
First, lossless-claw saves every raw message into a SQLite database, which means the original conversation is never discarded. Next, it groups older messages and asks a language model to generate summaries. The key design decision is that those summaries are stored in a directed acyclic graph (DAG), so the system can keep rolling lower-level details into higher-level thematic summaries as the conversation grows.
When an agent needs a precise detail from earlier in the session, lossless-claw exposes retrieval tools such as lcm_expand_query. Those tools let the agent drill back down through the summary graph and recover the original source material instead of relying on a lossy recap.
Why You Need lossless-claw
lossless-claw is compelling because it improves both reliability and operating efficiency for long-running agent workflows.
- It delivers strong measured performance. In the OOLONG benchmark, a lossless-claw-powered system scored 74.8, ahead of Claude Code at 70.3.
- It removes the need to constantly restart conversations. Agents can keep working for days or weeks while context usage stays within a manageable range.
- It avoids the typical "brain damage" effect that follows aggressive lossy compaction, because exact past details remain recoverable.
How to Install and Configure lossless-claw
Getting started with lossless-claw is straightforward once your OpenClaw environment is ready.
1. Verify the prerequisites
Before installation, confirm that you are running Node.js 22 or later and that OpenClaw already has an LLM provider configured for summarization.
For the best search performance, use a Node runtime compiled with SQLite FTS5 support. Without FTS5, lossless-claw can still work, but full-text retrieval will fall back to slower search behavior.
2. Install the plugin
Use the built-in OpenClaw plugin manager:
openclaw plugins install @martian-engineering/lossless-claw
That command downloads the plugin, records it in the plugin registry, and enables lossless-claw for the current system.
3. Set lossless-claw as the default context engine
To make lossless-claw your default context manager, check the OpenClaw JSON configuration and ensure the contextEngine slot points to lossless-claw:
{
"plugins": {
"slots": {
"contextEngine": "lossless-claw"
}
}
}
Restart the OpenClaw gateway after updating the configuration so the new context engine is fully loaded.
4. Tune the advanced settings
You can refine lossless-claw behavior with environment variables. A practical baseline is:
LCM_FRESH_TAIL_COUNT=32to protect the 32 most recent messages from compaction.LCM_INCREMENTAL_MAX_DEPTH=-1to allow the summary tree to grow without an artificial depth limit.LCM_CONTEXT_THRESHOLD=0.75so compaction begins when the context reaches 75 percent of capacity.
Conclusion
lossless-claw changes long-context memory management from reactive forgetting to active preservation. By storing original messages, compressing them into a retrievable DAG, and letting agents recover exact history on demand, lossless-claw makes persistent OpenClaw sessions much more practical.
If you want an AI assistant that can keep working without constantly losing track of past decisions, lossless-claw is one of the most important OpenClaw plugins to install.
Latest from the blog
New research, comparisons, and workflow tips from the Vibe Coding Tools team.
WebMCP replaces brittle screenshot-based automation with structured browser-native tools, letting AI agents interact with websites directly, accurately, and securely.
Learn how Claude Code Agent Teams orchestrates multiple AI teammates with shared tasks and mailbox messaging, and how to launch and scale this workflow.
Today’s AI assistants are smart but lack memory and initiative—every chat resets context, and they never act unless you ask first. Moltbot changes that.
