Tiered storage
A storage architecture that assigns memories to hot, warm, or cold tiers based on recency, trust score, and retrieval frequency. Hot memories live in fully-indexed OpenSearch with HNSW acceleration; warm memories use standard indexing; cold memories are archived in object storage. Movement between tiers is automated.
Tiered storage is the cost-control architecture for memory at scale. Keeping every memory in high-performance HNSW-indexed OpenSearch is viable when the memory store is small; at millions of memories it becomes prohibitively expensive. The insight is that access patterns are highly skewed: recent, high-trust memories receive the vast majority of retrieval traffic, while older or low-trust memories are rarely needed and can tolerate higher retrieval latency. Hot tier memories (recent, high trust, frequently retrieved) get HNSW-accelerated vector search at full speed. Warm tier memories (important but less frequently accessed) use standard OpenSearch indexing - slower but still reasonable for non-critical-path queries. Cold tier memories (archived, low value, rarely retrieved) live in object storage, retrieved only when explicitly requested or during full corpus analysis. Tier assignment is automated based on recency, trust score trajectory, and retrieval frequency; memories that suddenly become relevant are promoted automatically.