Memory (AI)

Memory in AI systems refers to the mechanisms by which models and agents retain information beyond a single interaction. While language models have a context window that serves as short-term memory within a conversation, AI memory systems extend this to persist information across sessions and interactions. There are several types of AI memory. Short-term memory is the conversation context within a single session. Long-term memory stores information across sessions — user preferences, learned facts, past interactions. Episodic memory records specific past events and their outcomes. Semantic memory stores general knowledge and facts. Working memory holds information being actively used for the current task. Implementation approaches include conversation history storage and retrieval, vector-based memory (embedding past interactions for semantic search), structured memory stores (databases of facts and preferences), and summarisation-based memory (condensing long histories into summaries).

Memory transforms AI from a stateless tool into a personalised assistant that improves over time. A customer support AI with memory remembers a customer's previous issues, preferences, and account details. A research assistant with memory builds on previous findings rather than starting fresh each time. For businesses, memory enables key capabilities: personalisation (adapting responses to individual users), continuity (picking up where previous interactions left off), learning (improving based on past successes and failures), and efficiency (avoiding redundant information gathering). Implementing memory requires careful consideration of privacy and data governance. What information should be remembered? How long should it persist? Who has access to the memory? How are GDPR deletion rights handled? These questions must be addressed as part of any memory system design.