A memory policy is the set of rules that govern what an AI agent stores in memory, what it discards, how it compresses older information, and how it prioritises what to retrieve. It is the decision-making logic that turns a raw store of past events into a useful, manageable memory.
Policies are necessary because an agent cannot keep and surface everything. Without rules, memory grows without bound, retrieval becomes noisy, and costs balloon. A memory policy decides questions like: should this interaction be remembered at all? Should several old conversations be summarised into one note? When memory is full, what gets forgotten first — the oldest items, the least important, or the least recently used? How much weight should recency carry against relevance?
These policies are what distinguish a genuine memory system from a plain vector store. The database handles similarity retrieval, but the memory policy supplies judgement about importance, recency, and compression. Well-designed policies let an agent remember what matters, forget what does not, and stay coherent and affordable as its history grows ever longer.