Long-term memory for AI
How AI could remember useful information across conversations over time. An active research area.
Forward-looking: this describes work in progress, not a finished product. Nothing here is a promised date or release.
What is it?
Most language models do not remember past conversations on their own. They only see what fits in their context window, the text provided in the current session. Long-term memory research looks for ways to let AI keep and recall useful information over weeks or months.
Current approaches include saving notes or summaries, storing past content in a searchable database and retrieving relevant pieces later, and making context windows much longer. Open problems include deciding what to remember, forgetting outdated or private information, avoiding wrong memories and giving users control.
Which models use it?
- Some chat assistants, such as ChatGPT, Claude and Gemini, offer memory features that save details users choose or allow.
- Retrieval-augmented generation (RAG) is a common way to give models access to stored information.
- Models with very long context windows can read large documents in one session.
- Research projects such as MemGPT explore managing memory for language model agents.
How to use it
- Check whether your AI assistant has a memory feature and what it stores.
- Review, edit and delete saved memories regularly.
- For projects, keep a short summary of key facts and paste it at the start of new chats.
- If you build apps, try a simple retrieval setup that stores notes and searches them when needed.
- Follow research on memory in AI agents, noting how memory is evaluated.
Benefits
- Could make assistants more helpful by remembering preferences and past work.
- Reduces the need to repeat the same background information.
- Helps agents work on long, multi-session tasks.
Limits and risks
- Stored memories can be wrong, outdated or taken out of context.
- Memory raises privacy concerns about what is kept and who can see it.
- Retrieval can miss important information or pull in irrelevant details.
- Reliable, controllable long-term memory is still an open research problem.
Prompt template
Copy it, replace the parts in [brackets] and paste it into any AI chat.
Summarise our project so far into a short 'memory note' I can paste into a future chat. Include: goal, key decisions, important facts, open questions and next steps, in under [150] words. Leave out any personal or sensitive details. Project notes: [paste your notes or conversation]
More in Research frontier
- World modelsResearch on AI that builds an internal model of how the world works. Early and uncertain.
- Continual learningResearch on AI that keeps learning new things without forgetting old ones.
- Interpretability (looking inside models)Research that tries to understand what happens inside AI models and why they answer as they do.
- Multimodal reasoningResearch on AI that reasons across text, images, audio and video together.
- Efficient small modelsMaking AI models smaller, cheaper and faster while keeping them useful.
- Alignment researchResearch on making AI systems reliably follow human intentions and values.