What is OpenClaw ContextEngine?

ContextEngine is a pluggable architecture introduced in v2026.3.7 with 7 lifecycle hooks (bootstrap, ingest, assemble, compact, afterTurn, prepareSubagentSpawn, onSubagentEnded) that lets developers replace OpenClaw's default context management strategy.

What is lossless-claw?

Lossless-claw is the first ContextEngine plugin that replaces sliding-window compaction with a DAG-based summarization system. Every message is persisted to SQLite — nothing is ever deleted. Agents can search and expand compressed history on demand.

How does lossless-claw perform on benchmarks?

On the OOLONG benchmark, lossless-claw scored 74.8 vs Claude Code's 70.3 using the same model. The gap widens with longer contexts.

What was wrong with OpenClaw's old context management?

Previous versions used hardcoded, lossy context compaction. When conversations hit token limits, the system discarded old messages or generated rough summaries, losing critical details and temporal ordering.

How is ContextEngine different from hooks?

ContextEngine is a slot, not a hook. Hooks are additive (multiple plugins can listen). Slots are exclusive — only one ContextEngine runs at a time, ensuring consistent context management strategy.

What other plugins use ContextEngine?

Beyond lossless-claw, there are plugins for persistent cross-session memory (MemOS Cloud integration), compliance controls (redaction before model sees content), and model-specific context tuning.

← Back to dashboard

clawsmith.com/signal/lossless-claw-contextengine-pluggable-memory-v2026-3-7

📈 TrendsUnknownFrameworkLive

Lossless-Claw Plugin + ContextEngine: OpenClaw v2026.3.7 Makes Memory Pluggable — Beats Claude Code on OOLONG Benchmark

v2026.3.7 introduces pluggable ContextEngine with 7 lifecycle hooks. The lossless-claw plugin replaces sliding-window compaction with DAG-based summarization, scoring 74.8 on OOLONG benchmark vs Claude Code's 70.3. 89 commits, 200+ bug fixes. Community waited 6 months for this.

Product Idea from this Signal

A memory system that gives OpenClaw agents persistent recall that survives session crashes and context window degradation

49.8k ▲

OpenClaw agents lose their entire conversation history between sessions even when the files exist on disk. Silent daily session resets wipe agent memory without warning. Meanwhile, every frontier LLM degrades past 50K tokens (proven by Chroma across 18 models), meaning even within a session, agents progressively forget earlier context. ByteDance's OpenViking (19K stars in two weeks) proves massive demand for agent memory infrastructure. This tool gives OpenClaw agents a persistent, queryable memory layer that survives crashes, session boundaries, and context window limits by storing structured knowledge externally and injecting only relevant memories per turn.

DEVTOOLCLIAI-AGENTOPEN-SOURCE

CompetitiveView Opportunity →