The Personal Small Model (PSM): Memory as a Learned Cognitive Primitive for Large Language Model Agents

We propose the Personal Small Model (PSM), a novel architecture for AI agent memory in
which a small, per-deployment model is trained not to store user content, but to master memory
operations: consolidation, decay scheduling, recall weighting, interference detection, and sleep-time
reorganization. Unlike existing approaches that treat memory as a retrieval problem—injecting
database fragments into a language model’s context—the PSM treats memory as a learned cognitive
skill, architecturally separated from the primary reasoning system. The PSM’s weights remain
shared and stable across all users; personalization lives entirely in per-user memory stores that the
PSM manages. This design eliminates catastrophic forgetting by construction, enables biologicallyinspired
memory consolidation, and allows a large language model to benefit from rich personal
context without any modification to its architecture. We present the full system design, training
methodology, memory tier hierarchy, and a sleep-time consolidation algorithm. This document
constitutes a public prior art disclosure. No patent is sought.