Access-Weighted Memory Decay and Reconsolidation in a Persistent Embodied Agent

Most AI assistants treat memory as a context window: a fixed buffer populated at session start and discarded at session end. This paper describes a sliding window memory decay model that treats importance as a function of recency and access frequency rather than a static property, implemented in Potato, a local AI agent running on a MacBook Air M4. Three primary contributions are described: (1) a ten-step sliding window importance decay curve, original to this work, that produces vivid recent memory and fades older memories to background retrieval; (2) a reconsolidation mechanism that reinforces memories upon retrieval with diminishing returns for repeated access; and (3) a nightly sleep decay system that models the flashbulb effect, where frequently accessed memories resist forgetting proportional to recency and access count. Together these produce a memory system where what the agent remembers is shaped by what the user actually engages with, not arbitrary retention schedules.