Limited memory AI is a label for systems that use recent observations to make a decision but do not keep those observations as lasting experience. The term comes from a four-part classification of AI that Arend Hintze of Michigan State University laid out in The Conversation on November 14, 2016, where the example was a self-driving car tracking the speed and direction of nearby vehicles. A rough everyday parallel, offered here as an analogy, is a chatbot's context window, the text a language model can refer to while writing a reply.
Key takeaways
- Hintze's scheme has four types: reactive machines, limited memory, theory of mind, and self-awareness.
- In the limited memory type, observations are transient and are not added to a store of experience the machine learns from.
- A language model's context window is its working memory for one conversation, and it is separate from the data the model was trained on.
- According to Anthropic's documentation, accuracy and recall can degrade as the amount of text in the window grows.
- Chat interfaces may drop the oldest text first when the window fills, so early instructions can fall out of a long session.
Where the label comes from
Hintze described four categories in 2016. The first, reactive machines, have no memory of the past and respond only to what is in front of them; his example was IBM's Deep Blue chess computer. The second, limited memory, can look into the past. The third, theory of mind, would model other people or creatures as having thoughts and emotions that shape behavior. The fourth, self-awareness, would form representations of itself, and he called it the last step and still distant. The scheme is an explanatory framework from one article and not a legal or technical standard.
The self-driving car example
Hintze's description of the second type is specific. A self-driving car observes other vehicles' speed and direction, and those observations are added to a preprogrammed picture of the road that includes lane markings, traffic lights, and curves. The car uses the combined information to decide when to change lanes safely. He stresses that the observations are only transient: they are not saved to the car's library of experience the way a human driver accumulates experience over years. The distinction is between using recent information and learning from it permanently.
Context windows: a working memory for chatbots
Anthropic's documentation on context windows defines the term as all the text a language model can reference when generating a response, including the response itself. It describes this as a working memory and says it differs from the large body of data the model was trained on. As a conversation advances, each user message and each reply accumulates in the window. The documentation lists capacities of up to one million tokens on some models, and notes that more context is not automatically better: as the token count grows, accuracy and recall can degrade, a phenomenon it calls context rot.
The same page notes that chat interfaces can manage the window on a rolling first-in, first-out basis. If a request's input alone exceeds the window, the programming interface returns an error. These details concern one vendor's documentation, and other products handle long chats in their own ways.
Four types side by side
| Type | Memory of the past | Example in the 2016 article |
|---|---|---|
| 1, reactive machine | None | Deep Blue, IBM's chess computer |
| 2, limited memory | Recent observations, transient | Self-driving car tracking nearby vehicles |
| 3, theory of mind | Would model others' thoughts and emotions | Described as a future type |
| 4, self-awareness | Would represent its own internal states | Described as the final step, still far off |
A classroom experiment on chatbot memory
A short test makes the idea concrete. The steps below are an illustrative exercise, and results will differ by product.
- Open a new chat and state a made-up fact, such as an invented class mascot name.
- Ask about it immediately, then again after ten unrelated questions, then after twenty or more.
- Note whether the answer stays correct, changes, or disappears.
- Repeat with a long pasted passage and compare.
- Record the product, the date, and the number of messages, since behavior varies and changes with updates.
The point is to see that recall inside a session is bounded and is not the same as learning. Whatever a student types is also data, so the exercise should use invented facts and no personal information.
What follows for teaching and study
A tutoring chat that runs for an hour may lose track of rules set at the start, and a long reading pasted early may be handled less reliably than a short one. Restating the key instruction before a new task, and splitting large documents into sections, are habits consistent with the documentation's point that curating what sits in context matters as much as how much fits. Facts that come back from the tool still need checking against a textbook or primary source.
FAQ
Is limited memory AI the same as machine learning?
No. Hintze's types describe how a system uses information over time, while machine learning describes how systems are built from data. A limited memory system may or may not use machine learning.
Does a chatbot learn from a conversation?
Within one conversation it can refer to earlier turns because they sit in the context window. That is different from the model being retrained. Products may also offer separate saved-memory features, which vary by vendor.
Why does a chatbot forget instructions in a long chat?
The window has a fixed capacity, some interfaces drop the oldest text first, and the documentation says accuracy and recall can degrade as the window fills. Restating the instruction usually helps.
Is a self-driving car limited memory AI?
Hintze used it as his example of limited memory in 2016. Real vehicles are complex systems, and the label is a simplification for teaching rather than an engineering category.
Related reading
- Limited memory AI and its constraints
- Everyday examples of limited memory AI
- How AI differs from machine learning
- AI and learning in schools
More in AI in education.
Last verified: October 2026
Sources
- Understanding the four types of AI, from reactive robots to self-aware beings, Arend Hintze, The Conversation, November 14, 2016
- Context windows, Claude API documentation, Anthropic, 2026