What is a Long-Term Memory (AI)?

A language model itself has no memory beyond the context window of the current conversation; anything it "remembers" across sessions has to be explicitly stored and retrieved by the surrounding system. Long-term memory implementations typically extract salient facts from a conversation (a user's stated preference, a decision that was made, a fact that was corrected), store them in a structured or vector-searchable store, and retrieve the relevant subset back into context when a new, related conversation begins. The central design challenge is relevance and staleness: naively storing everything and retrieving by similarity search surfaces outdated or contradictory memories just as readily as current ones, so mature memory systems track when a fact was learned, let newer information supersede older contradicting entries, and prune memories that are no longer relevant to the user's current context. Getting this wrong produces an agent that confidently repeats something the user corrected weeks ago.