How Much Storage Does Agent Memory Require

What a Single Memory Costs

Each memory entry has three storage components. The text content is the smallest: a typical observation of 50 to 200 words is 200 to 800 bytes. The metadata (timestamps, confidence scores, entity tags, agent ID, source context) adds 200 to 500 bytes as JSON. The embedding vector is the largest: a 1536-dimension float32 vector (OpenAI's text-embedding-3-small) consumes 6,144 bytes, and a 3072-dimension vector (text-embedding-3-large) consumes 12,288 bytes.

The total per memory ranges from 6.5 KB (small text, minimal metadata, 1536-dim embedding) to 14 KB (longer text, rich metadata, 3072-dim embedding). Most production systems use 1536-dimension embeddings, placing the typical memory at 7 to 10 KB.

How Storage Scales

Storage growth depends on three factors: how many memories the agent stores per session, how many sessions run per day, and how many agents operate in the system.

A conservative agent (stores 5 to 10 memories per session, 1 session per day) accumulates 50 to 100 KB per day. A moderate agent (10 to 20 memories per session, 3 to 5 sessions per day) accumulates 300 to 1000 KB per day. An aggressive agent (20 to 50 memories per session, 10+ sessions per day) accumulates 2 to 5 MB per day.

For a team of 10 moderate agents, monthly storage before lifecycle management is roughly 100 to 300 MB. This is well within the capacity of any modern database. Storage is rarely the bottleneck for agent memory systems; retrieval performance (searching through the stored memories efficiently) is what requires attention at scale.

The Knowledge Graph Adds Overhead

If the memory system maintains a knowledge graph (entity nodes and relationship edges), this adds storage proportional to the number of unique entities and relationships. Each entity node is small (100 to 500 bytes for the entity name, type, and metadata). Each relationship edge is similarly small (100 to 300 bytes). A memory store with 10,000 memories might have 2,000 to 5,000 unique entities and 3,000 to 10,000 relationships, adding 2 to 5 MB of graph storage, or about 10 to 20% overhead on top of the memory content itself.

Lifecycle Management Keeps Storage Bounded

Without lifecycle management, storage grows linearly with usage. With consolidation, decay, and forgetting, the active memory store reaches a steady state where new memories are added at roughly the same rate that old memories are consolidated or removed. In practice, systems with active lifecycle management maintain an active store that is 30 to 50% the size of the total ever stored.

Consolidation is the biggest driver of storage efficiency. When the system merges 5 related memories about a topic into a single consolidated summary, storage for that topic drops by 80%. When stale memories that have not been accessed and have low confidence are removed, they stop consuming both storage and retrieval resources.

Adaptive Recall's pricing is based on memory count rather than raw storage, so consolidation directly reduces costs. The free tier includes 500 memories, which is sufficient for a single agent running moderate sessions for several weeks. The lifecycle management features (consolidation, decay, forgetting) are included at all tiers, keeping the active memory count bounded without manual cleanup.