lossless-claw: How a 7-Star Project Solves the Hardest Problem in AI Agents — Forgetting
Long conversations kill agents.
Three hours into a coding session, the agent forgets the architecture decision you agreed on. Yesterday’s file structure — gone. Last week’s deployment strategy — truncated out of the context window. Every AI coding agent hits this wall. lossless-claw solves it by compressing context without losing anything.
Source: lossless-claw — Martian Engineering
7 Stars, Maximum Technical Depth
The gap between star count and technical sophistication is enormous. 7 stars, 0 forks — yet 63 commits, 3 releases, a dual-language architecture (TypeScript 66.8% + Go 32.5%), and a complete terminal UI for database inspection and repair. This is an implementation of the LCM (Lossless Context Management) paper published by Voltropy.
The project is a plugin for OpenClaw that replaces sliding-window compaction — the standard approach of simply discarding older messages — with a DAG-based summarization system. The core promise: “Nothing is lost. Raw messages stay in the database.”
Depth-Aware Prompts Are the Real Innovation
The system builds a directed acyclic graph of summaries. Raw messages become leaf summaries (depth 0), which condense into depth 1, then depth 2, and so on. The critical design choice: each depth level uses a completely different prompt strategy.
Depth 0 (Leaf) preserves timestamps, file operations, and decisions in a narrative summary. Depth 1 creates chronological session summaries with deduplication. Depth 2 focuses on arcs — goals, outcomes, what carries forward. Depth 3+ retains only durable context: key decisions, relationships, and lessons learned.
This mirrors how human memory works. Recent events are recalled in detail. Yesterday is a summary. Last month is conclusions. Last year is lessons. Each depth’s prompt is engineered to match that layer of abstraction.
Every Summary Ends With an Escape Hatch
All summaries include an “Expand for details about:” footer listing what was compressed. When the agent needs specifics, it calls lcm_expand_query, which delegates a sub-agent to walk the DAG back down to the source messages and compose a precise answer.
Four agent tools enable memory exploration: lcm_grep (full-text and regex search), lcm_describe (summary/file detail lookup), lcm_expand_query (sub-agent deep recall), and lcm_expand (low-level DAG traversal for sub-agents only).
The design insight: giving agents tools to explore their past is more effective than stuffing everything into context. Humans don’t hold all memories simultaneously either — they recall on demand.
Not Production-Ready Yet, But the Architecture Matters
The project currently requires an unmerged OpenClaw PR (#22201). The community is tiny. Production use is premature. But the architecture — depth-aware prompt strategies, DAG-based summarization, sub-agent expansion — represents the most sophisticated open-source approach to the long-context agent problem.
Context windows keep growing, but Lost in the Middle degradation and cost remain. “Just use a bigger context” isn’t the answer. Intelligent compression is.
It feels like talking to an agent that never forgets. Because it doesn’t.
Comments 0