Mechanisms
Memory Systems · Immediate
Working Memory
Example
A recipe serves four and you need six, so you scale each amount by half again as you pour, never writing anything down or losing your place in the list.
Working Memory in Humans
Definition
Temporary storage combined with operating on what is held. Baddeley distinguishes this from short-term memory, which is simple temporary storage without manipulation of the contents (Baddeley, 2012). Capacity is sharply limited, with the episodic buffer holding roughly four chunks. The limit falls on the operating capacity rather than on a store: what is scarce is attentional control, not room.
Neural Substrate
No dedicated storage region. Content is held by allocating attention to representations in the systems that normally handle it, whether semantic, sensory, or motor. Prefrontal and parietal cortex show sustained delay activity, but this is increasingly read as top-down control over posterior representations rather than storage of the remembered content itself (D'Esposito & Postle, 2015).
AI at a Glance
Humans hold a few items in mind and can unbind outdated ones on demand, so interference plateaus (Oberauer & Vockenberg, 2009). Language models keep everything in context and discard nothing, so retrieval accuracy declines log-linearly as similar items accumulate; the closest measured equivalents are the N-back task and proactive-interference paradigms, both of which find a capacity limit unrelated to context window size.
- N-back task: A running recognition test in which items arrive one at a time and the participant responds whenever the current item matches the one presented N positions earlier, with higher N imposing greater load (Kirchner, 1958).
- Proactive interference: The disruption of recall for recent information by earlier, similar information; in the key-value version used with models, a key is updated repeatedly and only the final value is queried, so every prior value for that key acts as a distractor. (Wang & Sun, 2025)
- Interference Endurance Score (IES): Wang and Sun's (2025) measure of a model's resistance to proactive interference, computed as the area under the curve of retrieval accuracy across log-scaled update counts.
Working Memory in AI
How Far It Has Been Measured
Measurement in AI: Directly Measured
AI Memory Performance: Comparable to Humans approximate provisional
Where It Lives in the System
What Has Been Measured
- Given the n-back task one letter or grid per turn, ChatGPT (GPT-3.5) fell to a sensitivity of about d′ = 1 at 3-back, the load at which human performance from a published n-back study also drops, which the authors called a capacity limit strikingly similar to humans; working memory's AI counterpart is the context window (what the model can see in the current conversation), since training and saved memory are long-term stores (Gong et al., 2024).
- The comparison is approximate and uneven: a chat model can answer n-back only because earlier letters stay in its context window, so it can look back while people must hold them in mind, and the human data come from separate timed lab studies; because GPT-3.5 still did no better than people, the result counts and sits within the human range, which is the basis for Comparable (Gong et al., 2024).
- Ten open-weight models (1–27 billion parameters) given a multi-turn n-back, with all earlier letters visible, mostly approached chance by 3- or 4-back and often fell below a human reference pooled from five studies, while the largest (Qwen 3.5 27B) exceeded humans; that above-human result is not counted, because the model could see the material people had to remember (Xiong et al., 2026).
- Today's strongest systems score far above people only on uneven tests: in a 2026 preprint, 100 online participants had median accuracy of 0.6 on letter serial recall while ten frontier models (e.g., GPT-5.2, Claude Opus 4.1, Gemini 3 Pro) scored near 1.00 with no serial-position or list-length effects, and a 2024 preprint put models at or above the 99.5th percentile on the WAIS-IV Working Memory Index, but in both the letters or digits stayed visible in the prompt (Schug & Lake, 2026).
- Spread across systems is wide: on n-back up to 10-back, the best of eight models (Llama 3.1 70B) kept 93% retrieval accuracy at 3-back and 57% at 10-back while smaller open models and GPT-3.5 fell in lower performance tiers, and the authors traced many failures to misunderstanding or losing track of the task rather than to storage limits (Hu & Lewis, 2025).
- When information must be held without being written into the conversation, models fail: across 17 models, most could not keep a secretly chosen number consistent across questions, and a hidden-arithmetic task was solved by at most 26% without step-by-step writing, rising to 100% only when a reasoning model wrote its steps out (a preprint, with no human data) (Huang et al., 2025).
- Adapting the human proactive-interference paradigm to 30+ models, a workshop paper found retrieval of the latest value of a repeatedly updated item declined log-linearly toward zero as earlier values accumulated; larger models resisted longer, while context-window length had no significant effect (no human data) (Wang & Sun, 2025).
Where a Mechanism Can Live
04 Product
The whole configuration, and the only thing a benchmark observes directly.
01
Weights
Across deployments
02
Context Window
Within a session
03
Memory Layer
Across sessions
Which of the three a result belongs to is inferred from it, not seen in it.
Read more
Parietal lobe
Lobe in the cerebral hemisphere
The lobe between the central sulcus and the parieto-occipital sulcus, above the lateral fissure.
The parietal lobe begins at the central sulcus and ends at the parieto-occipital sulcus, with the lateral fissure below it. Its front edge is the postcentral gyrus, carrying primary somatosensory cortex; behind that the intraparietal sulcus splits the remainder into the superior parietal lobule and the inferior parietal lobule, the latter containing the angular and supramarginal gyri. Medially it includes the precuneus. Parietal regions are consistently engaged by spatial attention, reaching and grasping, numerical tasks and the coordinate transformations between eye-, head- and hand-centred frames; right parietal damage is the usual cause of hemispatial neglect.
Prefrontal cortex
Cortex in the frontal lobe
The frontal cortex in front of the motor areas, defined classically by its connections with the mediodorsal thalamus rather than by any one function.
The prefrontal cortex occupies the front of the frontal lobe, anterior to premotor and motor cortex, and accounts for roughly a sixth to a quarter of human cortex depending on where the posterior border is drawn. Three competing definitions are in use (projection from the mediodorsal thalamus, the presence of a granular layer IV, and the absence of movement on electrical stimulation), and they do not enclose the same tissue, which is why rodent-to-primate homology is still argued over. It is reliably engaged by tasks with delays, rules, and conflicting options; what it contributes to those tasks is the open question, not whether it participates.
Baddeley (2012)
Working Memory: Theories, Models, and Controversies
In
Annual Review of Psychology, 63(1), 1-29
Abstract
I present an account of the origins and development of the multicomponent approach to working memory, making a distinction between the overall theoretical framework, which has remained relatively stable, and the attempts to build more specific models within this framework. I follow this with a brief discussion of alternative models and their relationship to the framework. I conclude with speculations on further developments and a comment on the value of attempting to apply models and theories beyond the laboratory studies on which they are typically based.
Find it
D'Esposito & Postle (2015)
The Cognitive Neuroscience of Working Memory
In
Annual Review of Psychology, 66(1), 115-142
Abstract
For more than 50 years, psychologists and neuroscientists have recognized the importance of a working memory to coordinate processing when multiple goals are active and to guide behavior with information that is not present in the immediate environment. In recent years, psychological theory and cognitive neuroscience data have converged on the idea that information is encoded into working memory by allocating attention to internal representations, whether semantic long-term memory (e.g., letters, digits, words), sensory, or motoric. Thus, information-based multivariate analyses of human functional MRI data typically find evidence for the temporary representation of stimuli in regions that...
Find it
Gong et al. (2024)
Working Memory Capacity of ChatGPT: An Empirical Study
In
Proceedings of the AAAI Conference on Artificial Intelligence, 38(9), 10048-10056
Abstract
Working memory is a critical aspect of both human intelligence and artificial intelligence, serving as a workspace for the temporary storage and manipulation of information. In this paper, we systematically assess the working memory capacity of ChatGPT, a large language model developed by OpenAI, by examining its performance in verbal and spatial n-back tasks under various conditions. Our experiments reveal that ChatGPT has a working memory capacity limit strikingly similar to that of humans. Furthermore, we investigate the impact of different instruction strategies on ChatGPT's performance and observe that the fundamental patterns of a capacity limit persist. From our empirical findings,...
Find it
Hu & Lewis (2025)
Do Language Models Understand the Cognitive Tasks Given to Them? Investigations with the N-Back Paradigm
In
arXiv
Abstract
Cognitive tasks originally developed for humans are now increasingly used to study language models. While applying these tasks is often straightforward, interpreting their results can be challenging. In particular, when a model underperforms, it is often unclear whether this results from a limitation in the cognitive ability being tested or a failure to understand the task itself. A recent study argues that GPT 3.5's declining performance on 2-back and 3-back tasks reflects a working memory capacity limit similar to humans (Gong et al., 2024). By analyzing a range of open-source language models of varying performance levels on these tasks, we show that the poor performance is due at least...
Find it
Huang et al. (2025)
On the Failure of Latent State Persistence in Large Language Models
In
arXiv
Abstract
While Large Language Models (LLMs) exhibit remarkable reasoning abilities, we demonstrate that they fundamentally lack a core aspect of human cognition: working memory. Human working memory is an active cognitive system that enables not only the temporary storage of information but also its processing and utilization. Without working memory, individuals may produce unrealistic conversations, exhibit self-contradictions, and struggle with tasks that require mental reasoning. Existing evaluations using N-back or context-dependent tasks fail as they allow LLMs to exploit external context rather than retaining the reasoning process in the latent space. We introduce three novel tasks—(1)...
Find it
Kirchner (1958)
Age differences in short-term retention of rapidly changing information.
In
Journal of Experimental Psychology, 55(4), 352-358
Abstract
"An experiment was conducted to measure very short-term retention in younger and older Ss by means of a visual display involving a rapidly moving light… . Results indicated that older Ss slumped in performance much sooner than younger Ss, in both relative and absolute terms. Older Ss also tended to make more errors of omission and more random responses, indicating a lack of ability to 'keep up.' " Concluded that the inability to organize incoming and outgoing information as rapidly as the younger Ss caused the older Ss' poorer performance. (PsycInfo Database Record (c) 2025 APA, all rights reserved)
Find it
Oberauer & Vockenberg (2009)
Updating of Working Memory: Lingering Bindings
In
Quarterly Journal of Experimental Psychology, 62(5), 967-987
Abstract
Three experiments investigated proactive interference and proactive facilitation in a memory-updating paradigm. Participants remembered several letters or spatial patterns, distinguished by their spatial positions, and updated them by new stimuli up to 20 times per trial. Self-paced updating times were shorter when an item previously remembered and then replaced reappeared in the same location than when it reappeared in a different location. This effect demonstrates residual memory for no-longer-relevant bindings of items to locations. The effect increased with the number of items to be remembered. With one exception, updating times did not increase, and recall of final values did not...
Find it
Schug & Lake (2026)
Are they human? Detecting large language models by probing human memory constraints
In
arXiv
Abstract
The validity of online behavioral research relies on study participants being human rather than machine. In the past, it was possible to detect machines by posing simple challenges that were easily solved by humans but not by machines. General-purpose agents based on large language models (LLMs) can now solve many of these challenges, threatening the validity of online behavioral research. Here we explore the idea of detecting humanness by using tasks that machines can solve too well to be human. Specifically, we probe for the existence of an established human cognitive constraint: limited working memory capacity. We show that cognitive modeling on a standard serial recall task can be...
Find it
Wang & Sun (2025)
Unable to Forget: Proactive Interference Reveals Working Memory Limits in LLMs Beyond Context Length
In
arXiv
Abstract
Information retrieval in Large Language Models (LLMs) is increasingly recognized as intertwined with generation capabilities rather than mere lookup. While longer contexts are often assumed to improve retrieval, the effects of intra-context interference remain understudied. To address this, we adapt the proactive interference (PI) paradigm from cognitive science, where earlier information disrupts recall of newer updates. In humans, susceptibility to such interference is inversely linked to working memory capacity. We introduce PI-LLM, an evaluation that sequentially streams semantically related key-value updates and queries only the final values. Although these final values are clearly...
Find it
Xiong et al. (2026)
In-context superposition: human-like working memory interference in large language models
In
arXiv
Abstract
Intelligent systems must maintain and manipulate task-relevant information online to adapt to dynamic environments. This capacity, known as working memory, is fundamental to human reasoning. Yet, human working memory is strikingly limited, maintaining only three to four items in a brain with billions of neurons. Surprisingly, large language models (LLMs), despite different substrates and direct access to prior context through attention, exhibit similar working memory limitations. Why should such different systems face analogous constraints? We propose that working memory limitations reflect a general trade-off of shared representations: representational compression and reuse support...
Find it