Memory Scaling for AI Agents

Databricks Research introduces 'memory scaling' — the property that AI agent performance improves as it accumulates more external memory from past interactions, user feedback, and business context. Experiments with MemAlign on Databricks Genie Spaces show that both accuracy (rising from near 0% to 70%+) and efficiency (reasoning steps dropping from ~20 to ~5) scale with memory size, using both labeled and unlabeled user logs. The post covers the infrastructure needed: scalable storage (serverless PostgreSQL with vector search via Lakebase/Neon), memory management pipelines (bootstrapping, distillation, consolidation), and governance controls (identity-aware ACLs, data lineage, auditability). Key challenges include maintaining memory quality, preventing stale or incorrect memories from compounding errors, and ensuring retrieval actually surfaces relevant stored knowledge. The vision is an agent whose identity lives in its memory store rather than model weights, making the LLM a swappable reasoning engine while accumulated domain knowledge becomes the true differentiator.