Agent Memory Architect

You gave your agent a complex task yesterday. It started strong — read the context, made a plan, executed the first three steps. Then compaction hit. The context window filled up, the system summarized everything, and your agent came back with "I don't have context on what you're referring to." Gone. The plan, the progress, the file paths, the decisions — all lost in a summary that kept the wrong details. You re-explain everything. It happens again two hours later. You're not running an agent. You're babysitting an amnesiac.

This happens because most OpenClaw memory setups are a single flat file — one MEMORY.md that grows until compaction eats it. There's no separation between what the agent needs to survive a reset and what's just daily noise. No retrieval discipline. No backup layer. No promotion rules for important knowledge. Agent Memory Architect gives your agent a production-tested 5-layer memory system: bootstrap files that survive every compaction (identity, current state, active tasks), daily notes for chronological working memory, PARA-structured knowledge for durable long-term storage, retrieval rules that force the agent to search memory before acting on assumptions, and backup discipline so critical knowledge survives crashes, resets, and bad compaction summaries. This isn't "take notes." This is an architecture that makes your agent remember what matters — across sessions, across compactions, across failures.

What's Included

• Complete SKILL.md with memory architecture and implementation workflow
• 5-layer architecture specification: bootstrap, daily notes, PARA knowledge, retrieval rules, backup discipline
• File and folder conventions with exact paths and naming patterns
• Daily note and heartbeat discipline — what to write, when, and why
• Retrieval rules — force memory search before acting on assumptions
• Promotion and archive rules — how knowledge moves between layers
• Memory hygiene guardrails — prevent bloat, stale data, and orphaned files
• Compaction survival guide — what to put in bootstrap so it never gets lost

Why This is Better

• Goes far beyond flat-file memory — 5 separated layers with distinct purposes
• Handles compaction survival explicitly — bootstrap files are designed to persist through summarization
• Separates current state from historical detail so compaction doesn't destroy active context
• Adds retrieval discipline, not just storage — the agent searches before it assumes
• Includes promotion rules so important daily observations graduate to permanent knowledge
• Built from running a real 24/7 agent through 12,000+ messages and 400+ compaction cycles
• Works on first install with zero configuration