Memory (in AI systems)

Product or application state persisted beyond one model request and later retrieved or applied to another run.

AI memory can include user-approved preferences, summaries, factual records, prior decisions, embeddings, or task state. The application stores and selects this material; the model sees only what the product places in the current context or uses through tools. Products differ in scope, automatic capture, retention, visibility, and whether memory applies across chats, projects, or accounts.

Builder example

Memory can reduce repeated setup and support continuity, but it can also preserve a wrong fact, leak context across scopes, or retain sensitive data longer than intended. Separate record types, attach provenance and timestamps, define update and deletion behavior, and let users inspect memory that affects their results.

You spend twenty minutes explaining your project to an AI. The next morning you open a new chat and it asks 'how can I help you today?' with zero recollection.

Save key project context, preferences, and decisions to a memory system that reloads in new sessions. Let users see and edit what is stored.

Common confusion: Memory is persisted state; the context window is request capacity. A product may have memory without loading all of it, or carry prior conversation state without calling the feature memory.