Memory Architecture Skill

To prevent agent memory context window exhaustion, organize memory into a 3-layer structure that separates durable knowledge, working context, and session data. This approach archives daily logs and prunes memory through a scripted compaction workflow.

The 3-layer memory system organizes data into Layer 1: Per-Message logs, Layer 2: Per-Agent workspace, and Layer 3: Global MEMORY.md. This separation keeps agent memory coherent across sessions by isolating session data from durable knowledge.

To set up per-agent memory management for OpenClaw workflows, place the 5-file structure (MEMORY.md, AGENTS.md, TOOLS.md, SOUL.md, HEARTBEAT.md) in each agent workspace. Run the compaction workflow scripts to maintain coherent memory across daily logs and rosters.

Yes, you can use this memory architecture to stop state leakage in multi-agent swarms by maintaining per-agent working memory. Separating workspace data ensures agents retain distinct roles without leaking state across the global memory context.

The best way to compact agent memory and prune session logs is using the provided scripted compaction workflow. It archives daily logs and prunes memory, enforcing the 5-file structure to keep memory coherent across sessions.

You need a memory compaction workflow for AI agents when facing context window exhaustion or identity drift. Compaction is required to archive daily logs and prune memory, ensuring long-term global memory management remains coherent across sessions.