Does AI Have Memory? Context Windows, Stored Memory and Training, Explained

Does AI Have Memory? Context Windows, Stored Memory and Training, Explained

Ask five people whether AI has memory and you will get five answers, because they are describing different mechanisms. One person means the assistant recalled their dog's name last Tuesday. Another means the chatbot lost the thread after twenty messages. A third means the model can produce facts it absorbed during training.

All three experiences are real. They come from different parts of the system.

This guide separates those layers, using published explainers, an academic article, an essay and a policy piece, and ends with a checklist of where the cited material converges — and where it fails to line up — plus a decision tree for your own situation. A note on framing, repeated here because it also governs the title and summary of this page: the three-layer split below is this guide's organising device for reading the material, not a framework lifted from any single source. The individual facts are attributed and linked at the point they are made.

The short answer

Some AI systems store and recall information across time. Others hold information only for the length of a conversation. The same word, memory, gets applied to both.

On the storing side, TechSee's glossary entry defines AI memory as the ability of artificial intelligence systems to store, retrieve and utilise information over time to enhance their performance and provide more accurate, contextually relevant responses. IBM's explainer on AI agent memory, written by Cole Stryker, uses a similar frame for agents specifically: the ability of an artificial intelligence system to store and recall past experiences to improve decision-making, perception and overall performance. Cognee's academy chapter, dated 2025-10-17, puts it in terms of what it changes — enabling machines to store, recall and apply past information turns static models into dynamic systems that learn from interaction.

Read that overlap carefully. These are three publicly published explainers whose wording happens to converge on store, retrieve, apply — not three disinterested parties agreeing. Each is published on the publisher's own site (ibm.com, techsee.com, cognee.ai), by companies whose business involves AI or AI-support tooling. Cognee publishes that chapter on the same domain where it markets its own product: its pricing page advertises a free-forever tier at $0 per month with one workspace and one million tokens included, no card required. TechSee's site likewise markets AI-supported customer-service tooling, describing an agent that can see damage, estimate the cost and help the customer there and then, and inviting prospective buyers to fill out a form for pricing and packages. Overlapping wording between vendor explainers is a useful starting definition, not corroboration.

On the not-storing side, a Medium post by Sean Warman (28 February 2025) reproduces a "sliding window" description: information isn't stored in a traditional sense; instead a window of recent context is used to generate a response, which allows the system to refer back to earlier comments and maintain coherence. Treat that particular passage for what it is — a model's own account of itself, quoted in an article, not vendor documentation. It is an illustration of the mechanism, not an audit of it.

One scope note before the layers. This guide explains mechanisms, not the behaviour of any named product. Memory settings, whether chat history is used, and how long anything is retained differ from one assistant to the next and change over time — so check your own assistant's current settings and documentation rather than assuming what is described here applies to it.

Layer 1: the context window (why it feels like memory, then stops)

This is the layer where an assistant follows the thread of a conversation and then, at some point, doesn't.

IBM's explainer places the in-session case inside a taxonomy rather than a mechanism. It points to the Cognitive Architectures for Language Agents (CoALA) paper from a team at Princeton University for the different types of memory, and summarises short-term memory (STM) as what enables an AI agent to remember recent inputs for immediate decision-making — its example being a chatbot that remembers previous messages within a session and so provides coherent responses instead of treating each user input in isolation. That is a description of the behaviour and its label, not of the storage arrangement behind it.

For a description of a mechanism, Warman's post is the one that speaks to it: the sliding-window passage quoted there says information isn't stored in a traditional sense, and that a response is generated from a window of recent context, which allows reference back to earlier comments. Keep the hedge attached — that passage is a model's self-description reproduced in an article, not vendor documentation, and it describes that one system.

The limitation is where Cognee's chapter picks up. It names the short-term constraint as what makes current LLMs "forgetful", and describes memory persistence as the contrasting capability — allowing systems to recall prior knowledge or interactions even after the window resets.

That gives you a one-way diagnostic clue, and it is worth being precise about which way it runs. If something you said in an earlier session comes back in a fresh one, some mechanism persisted it. The reverse does not hold: if it doesn't come back, that is not proof nothing was kept. Retrieval can fail, be scoped to certain kinds of information, or be applied selectively, and a product can retain conversation data without ever surfacing it to you. Non-recall tells you the information wasn't applied — nothing more.

Layer 2: stored, persistent memory

This is the layer most people mean when they say an AI remembers them.

Cognee's chapter describes a conversational AI memory system as one that stores dialogue history, user preferences or relevant facts in a retrievable format — and frames the payoff as continuity, personalisation and deeper understanding, in the way human memory supports reasoning and learning. TechSee's glossary describes the long-term side as systems, particularly those using large language models, storing information that can be accessed over long periods — remembering facts, processes and user preferences — and the retrieval side as recalling relevant information based on the current task or conversation context.

IBM's explainer draws the contrast at the architectural level: unlike traditional AI models that process each task independently, AI agents with memory can retain context, recognise patterns over time and adapt based on past interactions. Its worked example is a thermostat-style case — instead of reacting only to the current temperature, a system can store and analyse past data to make more intelligent decisions.

Storing more is not automatically better. IBM's explainer notes that optimised memory management helps ensure AI systems store only the most relevant information while maintaining low-latency processing for real-time applications — relevance and speed, not volume, are the design targets it names.

Layer 3: what the model learned during training

There is a third sense of remembering that has nothing to do with your conversation: what a model absorbed while being trained.

Warman's post relays an interview with Simon Prince in which he details the limitations of ChatGPT, and of AI in general. Two points travel from that relay: that the requirement is not just storing data but understanding context, intent and the nuances of human communication, which vary widely; and that careful management is needed to ensure a model does not forget previously learned information — a phenomenon named catastrophic forgetting — while maintaining performance. That is a blog post reporting an interview about deep-learning limitations, not a training-process reference.

Here is this guide's own framing of why that layer is easy to confuse with the others, offered as explanation rather than as anything a cited source says. What a model absorbed in training lives in its weights: it shapes what the model tends to produce, and it is not a list you can open, search or delete an entry from. A retrievable record of your conversation — the kind of store Cognee describes, holding dialogue history and preferences in a retrievable format — is a different object entirely. So a model stating a fact it learned in training is not the same event as it looking up something you told it, even though both look like "it remembered" from the outside. It also means the failure modes don't transfer: catastrophic forgetting is the model losing prior training, not your assistant dropping your name.

The training layer is also where your own use loops back. A Cambridge Core article in *Memory, Mind & Media* (© The Author(s), 2024) observes that virtual assistants, memory apps and chatbots build on the fragments of the past that have fed and trained large language models, and that further exchanges in turn train or guide AI systems to offer answers more attuned to the prompts they are fed.

The design question nobody agrees on: should it forget?

One might assume more memory is always better. A doc.cc essay pushes back on that instinct, asking what if human forgetting is not a bug but a feature. Its argument: our brains are not built to store everything, and evolution did not optimise them to store the past in high fidelity — it optimised us to survive the present. "Infinite memory", the essay argues, runs against the grain of what it means to be human. It then draws the contrast directly: where nature embraced forgetting as a survival strategy, we now engineer machines that retain everything — past prompts, preferences, corrections and confessions.

The Cambridge Core article argues the stake is larger than product design. It makes a case for a "third way of memory", to recognise how the entanglements between humans and machines both enable and endanger human agency in the making and remixing of individual and collective memory — including the growth of AI agents with increasing autonomy and, in its words, infinite potential to make, remake and repurpose individual and collective pasts, beyond human consent and control.

Note what these two are and aren't. They are arguments about design goals and about agency, not technical contradictions of the engineering explainers above. A system can be built with retrieval and with deliberate forgetting; these pieces argue about which default we should want and who controls it.

What persistent memory means for you

An agstudies.org publication on AI memory and its risks frames the user-facing version plainly: AI memory is the ability of these systems to store and recall information about you from your conversations. It observes that AI companies are rapidly developing the feature, arguing it makes tools more helpful and personalised.

Before taking the points below at face value, note what that publication is. It is an advocacy piece: its recommendations are addressed to state attorneys general, and it argues for AG-encouraged standards and regulation. Its framing of risks is shaped by that purpose, so weigh the workplace and portability observations as arguments for a policy position rather than as neutral survey findings.

Three points from it are worth carrying into your own decisions:

• Controls vary by product. It notes that some AI tools might require you to actively opt in to memory features, giving you more upfront control, citing Google's approach with Gemini, which it says also promotes tools to view and delete memories. That hedge is the source's own, and the description of Gemini here is as characterised by agstudies.org — treat it as a pointer to look for those controls yourself, not as a specification of them.
• Work context adds complexity. It flags that using AI with memory features in the workplace adds another layer of complexity, including tricky issues around intellectual property in workplace settings.
• Portability is an open policy question. It suggests attorneys general can encourage standards and regulations that let users take their AI-generated memories with them when switching services, to foster competition and user choice — which is a case for something to exist, not a description of something you can rely on today.

Checklist: what the cited material says, and how well supported it is

A compact synthesis of the material linked in this guide. The last column matters as much as the first: most rows are one publisher speaking.

Point	Stated by	Strength of support
Memory means storing, retrieving/recalling and applying information over time	TechSee, IBM, Cognee	Overlapping wording across three vendor-published explainers
Memory-equipped systems behave differently from ones that treat each task in isolation	IBM	Single source
In-session coherence can come from a moving context window rather than storage	Sean Warman	Single source, and a model's self-description quoted in a post
The short-term constraint is what makes current LLMs forgetful	Cognee	Single source, published by a company that markets its own AI memory product
Memory has named sub-types (short-term memory for immediate decisions, per CoALA)	IBM	Single source, citing the CoALA paper
Retaining only the most relevant information is a deliberate engineering goal	IBM	Single source
LLM-based systems can store information accessible over long periods	TechSee	Single source
Persistent memory about users raises control, workplace and portability questions	agstudies.org	Single source, advocacy piece
Retaining everything is contested as a design default	doc.cc	Single source, essay
Human–machine entanglement both enables and endangers agency over memory	Cambridge Core	Single source, academic, short excerpt

One question is where the cited material visibly fails to line up. Asked what an LLM-based system holds, TechSee says such systems can store vast amounts of information accessible over long periods; Cognee attributes current LLM forgetfulness to the short-term constraint and treats persistence as the separate capability that survives a window reset; and the passage quoted in Warman's post says information isn't stored in a traditional sense at all. Whether that is a genuine contradiction or three statements aimed at different layers — a bare model versus a product with a memory feature built around it — is not something these excerpts can settle, and the wording in each leaves the boundary of "the system" undefined. Either way the practical lesson is the same, and it is the one habit worth taking from this guide: ask which component a claim is about before accepting it.

Decision tree: what to do about memory in the tool you use

Start with your situation and follow it down.

1. Is your problem that it forgets mid-conversation?

That is the window layer. Cognee's chapter identifies the short-term constraint as the source of LLM forgetfulness, and distinguishes persistence as the thing that survives a reset. Practical move: restate the essential context in the current conversation, and check whether the product offers a persistent memory feature at all rather than assuming continuity.

2. Is your problem that it forgets between sessions?

You want the persistence layer Cognee describes — recall after the window resets — backed by the kind of retrievable store Cognee and TechSee describe (dialogue history, preferences, relevant facts). Practical move: look for an explicit memory setting, not a longer conversation.

3. Do you want it to remember, or do you want it not to?

If you want it to: IBM's framing is the value case — retained context, patterns recognised over time, adaptation from past interactions.

If you do not: agstudies.org notes that some tools might require an active opt-in, and points to Gemini as an example that, as it characterises it, also promotes tools to view and delete memories — a characterisation from that advocacy publication rather than from Google's own current documentation. Practical move: before entering anything sensitive, find your product's own memory settings and its view-and-delete surface, and read what it says about retention. If you cannot find one, assume the tool may retain what you type, and remember that not seeing information resurface is not evidence it was discarded.

4. Is this a work account?

agstudies.org specifically flags that workplace use adds another layer of complexity, including intellectual property questions in workplace settings. Practical move: check your employer's position before feeding proprietary work into a memory-enabled assistant.

5. Might you switch tools later?

The same publication frames data portability as something regulators could encourage — a goal it argues for, not a guarantee that exists. Practical move: do not treat accumulated AI memory as an asset you can move.

So: does AI have memory?

Some systems do, in the specific sense these explainers define: they store information, retrieve it against the current context, and use it to respond better. Other systems only appear to, because a window of recent conversation travels with each request. And a third kind of remembering — what a model absorbed in training — is a separate concern with its own failure mode in catastrophic forgetting.

The useful question is not whether AI has memory. It is which of these three layers you are relying on, and whether you can see and delete what the second one has kept.

Coverage limits

Three results ranking for this query are not fully represented above. The Cambridge Core academic article is quoted here only from a short captured excerpt rather than the full paper, so its wider argument and evidence are not summarised. A Reddit thread on AI memory being abandoned and a LinkedIn piece arguing that AI has no memory could not be retrieved at all, because both sites' robots.txt disallowed fetching them. Counter-arguments and academic framing from those pages are therefore outside the scope of this guide.

Sources

• Why Can't AI Remember Anything? — Sean Warman, Medium
• What Is AI Agent Memory? — IBM Think
• What is AI Memory? — TechSee glossary
• TechSee pricing
• To grow, we must forget… but now AI remembers everything — doc.cc
• Does Your AI Remember Too Much? Understanding AI Memory and Its Risks — agstudies.org
• What is AI Memory? — Cognee Academy, Chapter 1
• Cognee pricing
• AI and memory — Memory, Mind & Media, Cambridge Core