LLMs are like Memento's Leonard: no memory, only context
A cognitive neuroscientist explains why LLMs lack true memory and need workarounds.
In a viral post, a cognitive neuroscientist turned product manager draws a striking parallel between LLMs and Leonard Shelby from Christopher Nolan’s Memento. Like Leonard, who suffers from anterograde amnesia and uses Polaroids, notes, and tattoos to compensate for his inability to form new memories, LLMs rely on a context window to simulate recall. But this is not real memory—it’s a static snapshot that lacks the flexibility of human memory: dynamic updates, attention anchoring, truth validation, and integration across experiences.
The author argues that while LLMs excel at rote retrieval within a bounded context, they cannot learn from interactions over time, prioritize important details, or distinguish fact from hallucination. These limitations force developers to build external memory systems (RAG, vector databases, prompt caching) as workarounds. The post resonates because it frames a technical constraint in a human cognitive lens, reminding us that true memory in AI is far from solved.
- LLMs function like Leonard from Memento: no memory formation, just external aids (context window).
- Human memory dynamically updates, anchors attention, and validates truth; LLMs lack all three.
- Workarounds (RAG, vector stores, prompt caching) are needed to make LLMs useful beyond single sessions.
Why It Matters
Reminds professionals that LLMs’ memory limitations require deliberate system design, not just better prompts.