Memory Architecture System — Four-Tier Persistent Memory

Every OpenClaw operator hits the same wall: MEMORY.md grows until the agent loads 50K+ tokens of context on every turn, misses things that are clearly in there, and starts forgetting details from three weeks ago. The default single-file memory system was never designed for long-running agents.

This skill is the architecture that actually works. Four tiers: daily session logs for same-session continuity, weekly consolidated files (one file replaces seven, 85% token reduction), a curated long-term MEMORY.md kept intentionally small, and a local SQLite index with FTS5 full-text search and vector embeddings for semantic retrieval.

The SQLite tier is the key. Instead of loading your entire memory file every turn, the agent queries the index and receives only the relevant snippets. Token cost for memory reads drops from 50K/turn to under 2K/turn. At 100 turns per day at Sonnet pricing, that saves roughly $7/day.

Includes the consolidation prompt template, MEMORY.md seed template, context injection rules for AGENTS.md, retention policy, and a migration guide for existing flat MEMORY.md setups. Everything works on any Linux or macOS host with no external API dependencies.