Explanations of large language models often describe context windows as “memory,” implying that the AI remembers previous conversation, like a human recalling facts or experiences.
Charming, but misleading.
The Metaphor Problem
-
Memory implies conscious retention, recall, and understanding.
-
Reality: context windows are sliding buffers of token sequences. They do not store experiences or meaning; they constrain the next-step predictions.
-
Treating them as memory encourages the belief that LLMs can learn mid-conversation, reflect on past interactions, or hold intentions.
Why This Is Misleading
-
Anthropomorphises storage — buffers are treated like cognitive processes.
-
Obscures relational computation — what appears as remembering is merely the actualisation of token correlations in context.
-
Encourages overestimation of model capability — users imagine continuity of thought and understanding where there is none.
The “memory” metaphor conflates functional constraints with mental faculties.
Relational Ontology Footnote
From a relational perspective, context windows are structural actualisations of potential relational patterns. They do not retain meaning across instances; they simply instantiate constraints that shape the ongoing sequence. Memory, as a cognitive faculty, does not exist here — only pattern alignment in real time.
Closing Joke (Because Parody)
If LLMs truly remembered, every session would begin with:
“Ah yes, I recall our discussion about Schrödinger’s cat last Tuesday. Shall we continue, or do you prefer a recap?”