Glossary of Common AI Terms · Purpose and Execution of Memory

A

Adaptive retrieval: A system that decides case by case whether to consult an external store before answering, rather than always or never doing so (Mallen et al., 2023).; Why it matters A judgement about when to look something up, made by software rather than by the model.
Agent: A model given tools and a loop, so that it can take actions (searching, running code, writing a file) rather than only answering.; Why it matters Most published AI memory architectures are agent architectures, and the memory layer is usually built for one.
AI model: A specific prediction algorithm, built from a fixed set of training data. ChatGPT runs on GPT models, Claude on ones named Opus, Sonnet, and Haiku. A new version means new training, not a software update.; Why it matters The part that carries a version number, and the only part of a platform that is trained rather than built.

B

Benchmark: A fixed set of tasks with an agreed way of scoring them, so that different systems can be compared on the same footing.; Why it matters Much of Table 1 turns on whether one exists: a capacity with no benchmark cannot be scored, only noted.

C

Causal tracing: A method for finding where a model computes something: corrupt an input, restore selected internal activations, and observe which restorations recover the original output (Meng et al., 2022).; Why it matters The closest AI equivalent of lesion-and-restore in neuropsychology, which is what makes the two literatures comparable at all.
Chunk: Two different things, and both appear on this site. In AI, one slice of a longer document, cut so that it can be embedded, stored, and retrieved on its own. In psychology, a unit of meaningful material held in working memory, of which roughly four fit at once (Cowan, 2001).; Why it matters The collision is worth noticing. The AI sense is a storage decision an engineer made and can revise; the human sense is a capacity limit nobody chose.
Confabulation: Producing fluent, confident content that is not grounded in anything real. Commonly called hallucination.; Why it matters In humans a symptom; in these systems arguably a default operating mode.
Context: The body of text a system is given alongside a question: a document, a codebase, a conversation history (Lin et al., 2025).; Why it matters The material itself, as distinct from the context window, which is the room it has to fit in.
Context window: Everything the model can see at once: your message, the conversation so far, and anything the system has pulled in for it. When a conversation outgrows it, depending on the product, the oldest material is dropped, earlier turns are condensed into a summary, or the conversation cannot continue.; Why it matters The closest thing to working memory, and the one that is actually measured.

E

Embedding: A long list of numbers standing in for a piece of text, arranged so that texts about similar things sit near each other.; Why it matters How relevance is judged during retrieval: by proximity, not by meaning.
Event segmentation: Parsing continuous experience into discrete units at boundaries, which is what turns a stream into an episode that can be retrieved as a unit (Fountas et al., 2025).; Why it matters Human memory does this automatically. A memory layer cuts on turn or token counts instead, which is why its stored units are fragments rather than occasions.

F

Fact extraction: The write-time step in which a memory layer distils a durable statement out of a conversation and discards the exchange it came from.; Why it matters The point at which a system decides what about you is worth keeping, using a rule somebody else wrote.
Fine-tuning: Further training on a narrower body of material, after the main training is finished, to shift a model's default behaviour. RLHF is one form of it.; Why it matters Changes disposition, not recall.

H

Hallucination: Output that is fluent, confident, and wrong. Hallucination is not a malfunction, just ordinary behaviour making the wrong prediction. A system whose only task is to continue text plausibly has no separate step at which it checks whether the continuation is true.; Why it matters This site files it under confabulation, the word the clinical literature uses.

I

Inference: A single run of the model: the moment it produces an answer.; Why it matters Nothing is retained between one inference and the next unless something outside the model saves it.

K

Key projection: The transformation an attention layer applies to each stored token to produce the index an incoming query is matched against (Dinh et al., 2026).; Why it matters Retrieval inside the context window, as opposed to retrieval from a store outside it.

L

Large language model (LLM): The kind of AI model this site is about: one built by learning to predict the next piece of text across an enormous body of writing.; Why it matters Every claim here concerns this family of systems, not AI in general.

M

Many-shot jailbreaking: Prompting a model with hundreds of fabricated examples of a behaviour so that it produces that behaviour on a request it would otherwise refuse (Anil et al., 2024).; Why it matters Repeated exposure shifting a response, entirely within one context window and gone when the session ends.
Memory layer: A store kept outside the model, into which pieces of past conversation are written and from which a few are fetched back when a new message arrives. Mem0, A-MEM, and Letta are examples.; Why it matters Layer 03 of a platform, and what nearly everyone means by AI memory. It sits beside the model, never inside it.
Model editing: Directly modifying weights to change one stored fact without retraining the model (Meng et al., 2022).; Why it matters The nearest thing to writing into the model itself, and the exception to weights being fixed after release.
Multimodal: A model that accepts more than text, such as images, audio, or PDFs.; Why it matters Sets what can enter the context window at all, and so what a memory layer has to work with.

P

Parameters: The numbers that make up the equation. (So in the famous y = mx + b there are two parameters, m and b.) GPT-3 had 175 billion of them (Brown et al., 2020). Open-weight models still publish their counts, now into the hundreds of billions; the companies behind ChatGPT, Claude, and Gemini do not publish counts for their flagship models.; Why it matters Far fewer numbers than the text they were fitted to, so the text cannot be stored whole. Much of it survives as a compressed trace, which is why a model can recall a great deal but has no copy to check against.
Persistent memory: A feature that writes selected facts to storage so they can be pulled back into a later conversation. Usually a summary someone or something chose to keep.; Why it matters Not recollection, but a saved note that gets re-read.
Pretraining: Building an AI model by fitting it to a body of training data, the way a line of best fit is fitted to a set of dots. What comes out is the equation, not the original dots.; Why it matters The step that makes a model, and the reason it holds no copy of what it learned from.
Product level: The shipped consumer application, where memory behaviour is set by product design rather than by the model or by any one retrieval component.; Why it matters Layer 04. Where the defaults you did not pick are chosen, and why two products on the same model behave differently.
Prompt: What you type, together with everything handed to the model alongside it: the conversation so far, attached files, and the platform's own instructions.; Why it matters The only channel into the model. Anything a memory feature recalls arrives this way, as text.

R

Reinforcement learning from human feedback (RLHF): Further training in which people rate candidate answers and the model is tuned toward the ones they prefer.; Why it matters Shapes manner and refusals rather than what the model knows. Disposition, not recall.
Repetition penalty: A sampling setting that discounts any token already used in the generated text, making it less likely to be chosen again (Keskar et al., 2019).; Why it matters A response decrement produced by a setting rather than by learning, which is why it is not habituation.
Retrieval index: An external store of documents searched at question time, whose results are handed to the model as text rather than held in its weights (Guo et al., 2026).; Why it matters The store half of retrieval-augmented generation, and the thing a memory layer usually searches.
Retrieval-augmented generation (RAG): Searching a store of documents for passages that look relevant, then handing those to the model along with the question.; Why it matters The dominant architecture for long-term AI memory (Lewis et al., 2020).
Ripple effect: The set of related facts that ought to change as a consequence of editing one fact (Cohen et al., 2024).; Why it matters Edits propagate poorly. A fact can be changed in one place and left standing everywhere it implies something.

S

Scaffold: A procedure imposed around a model to improve its behaviour without retraining it, such as requiring it to keep a running list, or splitting work across several coordinated instances (Liu & Gabriel, 2026).; Why it matters Capability added outside the model. Nothing about the model changed, so nothing transfers when the scaffold is removed.
Session: One continuous conversation. Reopening an old chat brings its transcript back, but a new chat starts without it unless a memory feature carries something across, which some products now do automatically by drawing on past chats.; Why it matters The boundary across which human and AI memory diverge most sharply.
Source metadata: Fields attached to a stored memory item recording which turn it came from. Some systems keep them; others discard them when compressing (Chang & Chen, 2026).; Why it matters Without them a recalled fact arrives with no occasion attached, which is most of what separates this from episodic memory.
System prompt: Standing instructions the platform places in front of your conversation, setting tone, rules, and what the model will refuse.; Why it matters Part of the product layer rather than the model, and why one model can behave differently in two products.

T

Temperature: A setting controlling how strictly the model sticks to the most likely next word. Lower is more predictable, higher more varied.; Why it matters Why the same prompt can produce different answers twice over, with nothing about the model having changed.
Token: A chunk of text, very roughly a short word or part of one. Models read and write in tokens, not letters or sentences.; Why it matters Every limit in this table is really a limit in tokens.
Training data: The text an AI model was built from (books, Wikipedia, Reddit, news, code, and a great deal of the rest of the web). It is read once, while the model is being built, and is never consulted again while you type.; Why it matters Often mistaken for memory; it is closer to having read a great deal once.
Turn: One exchange: your message, and the answer that follows it. The model runs once per turn and re-reads everything it has been given each time, because nothing carries from one turn to the next on its own.; Why it matters The unit the context window is rebuilt at, and the boundary anything unwritten is lost across.

V

Vector database: A store built to answer "what is near this?" quickly across millions of embeddings.; Why it matters Where persistent memory usually physically lives.

W

Weights: Another word for a model's parameters, and the usual one in technical writing. Fixed when the model is released and identical for every user in the world.; Why it matters The layer people mean when they say the model itself. Nothing you do in a conversation edits it.