martian-engineering/lossless-claw

Lossless Claw is a plugin for OpenClaw, an AI agent platform, that changes how the agent handles long conversations. AI models can only hold a limited amount of text in their active memory at one time. When a conversation grows too long, the agent normally discards the oldest messages to make room for new ones. Lossless Claw replaces that discard behavior with a smarter system that keeps everything. Instead of throwing away old messages, the plugin stores them in a local SQLite database, a lightweight file-based database that stays on your machine. As older messages accumulate, the plugin automatically summarizes groups of them using the connected language model, producing a compressed version. Those summaries can in turn be summarized further, creating a tree-like structure of condensed history. When the agent needs to respond, it combines the recent raw messages with the relevant summaries to stay within the model's memory limit. Nothing from the original conversation is discarded. The raw messages stay in the database, and every summary links back to the source messages it was created from. The plugin also gives the agent three tools it can use itself: one to search the conversation history by keyword, one to get a description of a compressed chunk, and one to expand a summary back into its original detail. This means the agent can actively look up things from far back in the conversation rather than simply having no access to them. Installation is handled through OpenClaw's plugin system with a single command. Configuration options let you tune how aggressively the plugin compresses older messages, which model it uses for summarization, and how large a chunk of messages must grow before it is compressed. A set of slash commands inside OpenClaw let you check the plugin's status, back up the database, diagnose broken summaries, and rotate session transcripts. The plugin is written in TypeScript and requires Node.js 22 or later alongside a working OpenClaw setup.