Reinforced Synaptic Memory: memory for AI agents that is rebuilt by use and pruned by disuse

Memory for AI agents is commonly described as persistent: information written once and kept until someone removes it. This note defines a different kind, Reinforced Synaptic Memory, in which a memory is kept in proportion to its use, fades in its absence, and is changed by the observed effect of acting on it. The definition rests on two independent grounds. The first is biological precedent: in the brain nothing persists by itself, a memory lasts because recall rebuilds it, and fast learning and stable retention are kept in separate systems on separate timescales. The second is an argument from what a memory is: a claim about the world whose only evidence is the outcome of acting on it, so that an unused memory cannot be told from a wrong one. The model has three stores, a bounded working memory for the session or task in hand, a long-term picture of the person, and a set of reflexes that act without being recalled, with consolidation as the only path between them. It ranks the evidence that may change a memory, keeps a person's preference from overruling a fact, and covers a short session, a long-running agent and a chain of delegating agents with the same rules, so that a delegate runs as the person's own session rather than as a link in a relay of guesses. We state what the model is not, say what would falsify it, address the objections we consider strongest, trace the model to the author's earlier proposal of a knowledge metabolism and reduce that proposal to four properties, and describe a reference implementation for AI coding assistants.