Three-Tier Episodic Memory Architecture

A SQLite-based memory architecture for AI agents that stores conversational history across three tiers: hot (last 7 days, full conversation turns, under 10ms retrieval), warm (days 7-30, daily summaries and extracted facts, around 100ms retrieval), and cold (archived quarterly rollups, compressed on disk). No cloud dependency, no vector database.

The key feature is temporal validity on facts. Rather than storing bare facts, each is stored with valid_at and invalid_at timestamps. A fact like 'Sean is stressed about wallet rotation' is only true for a specific window. Stale facts that never expire degrade memory quality over time; this architecture handles that automatically.

The daily consolidation process moves full conversation turns from hot to daily summaries in warm, extracts temporal facts, and archives warm records to cold after 30 days.

What is included: the full architecture specification, SQLite schema for all three tiers with indexes, the temporal fact extraction prompt template, the AGENTS.md memory integration block, and the daily consolidation script.

Best for operators building agents that need to track evolving context over weeks and months, not just within a session.