AI memory and human memory use the same words for different mechanisms. A large language model keeps something like memory in three places. The weights learned in training correspond to abstract long-term memory. The context window and its key-value (KV) cache play the part of working memory. External stores such as vector databases and memory banks keep persistent information beyond the context window. The mapping helps, and it breaks in three places you can predict: weights do not take in new experience after training, retrieval does not rewrite what it retrieves, and nothing in a standard model forgets the way people do.
AI memory and human memory, component by component
The table below sets out the comparison the rest of this article argues. Each row stands on its own.
| AI component | Closest human counterpart | Where the analogy holds | Where it breaks |
|---|---|---|---|
| Model weights (parametric memory) | Abstract long-term memory, closest to semantic knowledge | Patterns from training are compressed into parameters and switch on during the forward pass with no lookup step | Parameters freeze when training ends, so the model knows nothing after its cutoff until it is retrained |
| Context window and KV cache | Working memory | Small and temporary, and lost when the session ends | Models often fail to use their full nominal length, and earlier turns are kept whole until something truncates them |
| External store (RAG, vector database, agent memory bank) | Long-term memory | Holds records across sessions and brings them back when a cue matches | Retrieval only activates stored content and does not reshape it the way reconsolidation does |
| Eviction, summarisation and decay policies | Adaptive forgetting | Low-value information is dropped to keep the working set relevant | The forgetting is an engineered rule, not a cognitive process |
For a ground-level walk through what type of memory an AI model uses and how a neural network stores it, start there. This page is about the comparison with the brain.
Does an AI model remember things the way the brain does?
No. A trained language model does not have a memory system in the usual sense. What it learned from its training data is stored as patterns spread across billions of numerical parameters. Liang and colleagues’ 2025 survey AI Meets Brain links this parametric memory to abstract long-term memory in human cognition. It also lists the costs: a time lag after the knowledge cutoff, hallucination (because knowledge is held as probability distributions rather than exact records), and a potential risk of catastrophic forgetting.
The deeper difference is that the model is stateless. Each prompt is processed on its own, and a plain LLM carries nothing over between calls unless a memory system is added around it. In the brain, memory shapes how a person thinks and acts now and later, instead of just storing and replaying content. When a memory-augmented chatbot “remembers” you, that continuity comes from the memory system added around the model.
How does a context window compare with human working memory?
Here the comparison has measured support. Anthropic’s documentation for Claude calls the context window all the text the model can reference while generating a response. It separates this from the training corpus and calls it a “working memory” for the model. Human working memory is small and short-lived. A neuroscience review puts it at about four chunks held for up to about 20 seconds, close to George Miller’s older estimate of seven plus or minus two items. Liang and colleagues give a range of 4 to 9 items and a window of no more than 15 to 20 seconds.
A 2023 empirical study ran ChatGPT through verbal and spatial n-back tasks and found its working memory capacity limit “strikingly similar” to the human one. That result needs some care. It measures performance on n-back tasks, not the size of the window. Nominal windows are large: a survey cites 128,000 tokens for GPT-4o, and Anthropic documents up to 1M tokens depending on the model.
A bigger window is not a better memory
A larger window is not automatically better. Liu and colleagues (2023) found that language models answer best when the relevant information sits at the start or end of a long context. Performance drops significantly when the information sits in the middle, even for models built for long contexts. One study confirmed primacy and recency effects across many tasks and models, the same serial-position biases documented in human psychology. It also found that careful prompting reduces them only inconsistently. Anthropic describes the general version as “context rot”: accuracy and recall degrade as the token count grows.
Measured agent workloads show the cost side of the problem. The 2026 paper The Missing Memory Hierarchy instrumented 857 production sessions and found that 21.8% of input tokens were structural waste: unused tool schemas, duplicated content and stale tool results, with stale results reprocessed at a median amplification of 84.4×. Its authors argue that the context window is not memory at all. They describe it as an L1 cache, small, fast and expensive, which the field treats as the whole memory system. In several of their sessions, removing stale content improved output as judged by evaluators, which fits the idea that irrelevant tokens dilute attention.
KV cache pressure: where the analogy becomes an engineering problem
Human working memory is limited by cognitive resources. A model’s working memory costs GPU memory per token. During autoregressive inference the KV cache stores a representation of every token processed so far, so it grows with sequence length. For large models the per-token KV footprint is on the order of megabytes. A smaller context per request therefore means more concurrent requests per GPU. The cache also grows and shrinks dynamically. If it is managed badly, fragmentation and duplication waste memory and limit batch size, and offloading the cache to host memory runs into CPU-GPU bandwidth limits.
Eviction schemes such as Heavy Hitters Oracle (H2O) keep a balance of recent tokens and the few that receive most of the attention, and SnapKV compresses the cache down to positions the attention heads keep returning to. StreamingLLM keeps a sliding window plus a small set of “attention sink” tokens, so most past state is overwritten continuously. At the application level, MemGPT issues a memory-pressure warning when the prompt passes a threshold such as 70% of the window. This lets the model save important facts before older messages are evicted and summarised. These are cache policies, not cognition.
How do AI systems keep long-term memory, and how does that compare with the hippocampus?
In the brain, the hippocampus is not a storage warehouse for long-term memory. It indexes memory traces. New information is first encoded across distributed neocortical regions, the signals converge in the hippocampus, and the hippocampus works as an index that reactivates those distributed traces. Over time, systems consolidation moves memories toward the neocortex. During rest and sleep, hippocampal replay reactivates recent experiences and fits them into existing knowledge structures. Consolidation also changes the memory as it moves it. Consolidation into the neocortex is theorised to lose finer temporal, spatial and contextual detail, shifting episodic memory toward gist-like semantic memory.
AI systems add long-term memory through external stores outside the weights. Lewis and colleagues (2020) introduced retrieval-augmented generation (RAG). It pairs a pretrained parametric model with a non-parametric memory: a dense vector index searched by a neural retriever. External memory of this kind can be updated without retraining and makes provenance easier to track, which addresses the time lag and hallucination problems of parametric memory. MemGPT takes the idea from operating systems. It treats the context window as main memory and external storage as disk, and it pages information between the two. HippoRAG borrows directly from hippocampal indexing theory. It combines LLMs, knowledge graphs and Personalized PageRank to imitate the separate roles of neocortex and hippocampus.
The parallel also shows up in computation. Whittington and colleagues (ICLR 2022) showed that transformers with recurrent position encodings reproduce the spatial representations of the hippocampal formation, including place and grid cells.
Retrieval is not reconsolidation
The analogy fails on what happens at the moment of recall. In humans, a reactivated memory enters a fragile state in which it can be modified or updated. Memory is a dynamic system that later encounters keep reshaping. Liang and colleagues are explicit: retrieval in current agent architectures is mainly selective activation of stored content driven by contextual cues, not a process that reshapes the trace. Human memory also allows conflicting interpretations to coexist and resolves them by merging old and new knowledge. A vector store returns the closest matches to a query by similarity, stale entries included. Keeping stale entries from biasing retrieval needs careful regulation of updates.
Is the human brain smarter than AI?
On memory specifically, the brain does several things current AI systems do poorly. A study drawing on human cognitive scenarios describes the brain refining knowledge progressively without erasure, balancing stability with plasticity. Parametric storage, by contrast, suffers catastrophic forgetting, and structured databases need manual updates. The same study finds that AI memory modules are kept apart from each other, with vector stores separate from knowledge graphs, while people tie sights, emotions and context into one memory. Humans also compress experience into general principles that apply across situations, which current LLMs do poorly.
AI systems have their own strengths. External memory modules have theoretically unbounded capacity and can retain records across sessions. Retrieval-based systems can show exactly where a fact came from. Human memory, by contrast, cannot keep original details through consolidation. On memory, the brain is stronger at adaptive memory and machines at exact retention.
Do AI systems forget information the way people do?
They do not, and forgetting is one of the most critical differences. Human memory has an adaptive forgetting mechanism. It prevents overload, helps the brain prioritise, and lets it drop outdated information. In the brain, forgetting actively filters out redundant information. The brain strengthens or prunes memories according to importance and frequency, and when working memory nears its limit it gives priority to task-relevant information and suppresses the rest. Consolidation then makes the memories that survive more resistant to decay and interference.
An LLM’s knowledge does not fade in this way. A trained model keeps everything its training gave it and changes only when retrained. Within a session, nothing is dropped from text, KV cache or hidden-state memory unless something intervenes, so information piles up, which can slow retrieval. AI systems show two kinds of apparent “forgetting”:
- Catastrophic forgetting. Weights work like synaptic long-term memory, but unlike human memory they are prone to catastrophic forgetting when trained on new data.
- Truncation. A model seems to forget early parts of a long conversation because of token limits, which the survey calls a technical constraint rather than a cognitive process.
Some agent systems design forgetting in. SAGE borrows the Ebbinghaus forgetting curve to drop low-value information adaptively, and MEMORYLLM keeps a fixed-size latent memory pool that balances old and new knowledge. The economics favour it. The Missing Memory Hierarchy points out that a 5,000-token file kept in context for 20 turns costs 100,000 input tokens of processing, against 5,000 to re-read it once when it is actually needed.
The brain makes its forgetting decisions without being asked. In an AI system, those decisions need explicit programming: which tool results expire, which records go stale, what the summary keeps and what it drops. A context window left to fill up carries the waste and attention dilution the measurements describe. Teams that write a forgetting policy are building the part of memory a standard model lacks.
FAQ
Is AI a human brain?
No. A language model stores patterns as numerical parameters and has no memory system in the biological sense. Some computations do line up: transformers with recurrent position encodings reproduce hippocampal place and grid cell representations, but that is a computational parallel, not an equivalence.
Is AI or humans better at remembering?
It depends on what you need to remember. Humans integrate, generalise and forget adaptively, while AI external stores keep exact records with traceable provenance.
