Memory Scaling for AI Agents

This post introduces memory scaling as a new design axis for AI agents, showing that performance improves as external memory grows.

• •Memory scaling is distinct from parametric and inference-time scaling, closing domain knowledge gaps neither can address alone
• •Two memory types are defined: episodic (raw interaction records) and semantic (distilled rules/patterns), each requiring different storage and retrieval strategies
• •MemAlign experiments on Databricks Genie Spaces show accuracy rising from near zero to 70% with labeled data, surpassing expert-curated baselines by ~5%
• •With only 62 unlabeled user logs, the agent exceeded expert baselines while reasoning steps dropped from ~19 to ~4.3
• •A pre-computed organizational knowledge store improved accuracy by ~10% on benchmarks requiring vocabulary bridging and table-level knowledge