Why LLMs Live in a World Like Leonard in Memento

In Christopher Nolan’s *Memento*, the protagonist, Leonard Shelby, suffers from anterograde amnesia: he cannot form new long-term memories. Every few minutes, his mind resets. To make sense of the world, he relies on notes, Polaroids, and tattoo, essentially fragments of externalized context he can trust when his memory resets.

Large Language Models (LLMs) live in a world much like Leonard’s.

They don’t actually remember. Each time you prompt one, it starts from zero. The model has no intrinsic awareness of what you said five minutes ago, unless you feed that text back in. What it does have is a body of statistical “tattoos”, the patterns it absorbed during training. That’s its world: a map of everything it’s seen before, frozen in time.

To help it function better, we build scaffolds around it.

• The context window is the short-term memory — a Polaroid of recent events.
• Retrieval-augmented generation (RAG) is the notebook — a stack of labeled fragments it can reread to stay consistent.
• System prompts act like Leonard’s tattoos: reminders of purpose, identity, and constraints.

But the similarity goes even deeper. Like Leonard, LLMs can’t easily tell the difference between real and fabricated memories. If a note (or a data source) is wrong, out of context, or summarized with loss of critical fidelity, it will still trust it. The illusion of coherence replaces true continuity. That’s why hallucinations happen: the model fills in blanks with whatever fits the pattern.

Leonard’s system was clever but fatally flawed. By trying to outsource memory to objects, Polaroids, notes, tattoos, he assumed he could reconstruct truth from static fragments. But without real memory, he lost the thread of causality. Each clue was decontextualized, stripped of how and why it mattered. He mistook the record for reality. Over time, that made his system manipulable: other people could plant false notes, or he could misinterpret his own. The evidence became circular and he trusted the marks on his skin more than the shifting facts around him.

The consequence was devastating. Leonard built a closed loop of belief, insulated from correction. Every time his system broke, he simply rewrote it, convinced he was chasing justice while actually chasing a fiction. His framework gave him certainty, but not truth. That’s the real cautionary parallel for LLMs. Without persistent, self-correcting memory , without the ability to verify and update context, they risk the same fate: coherence without understanding, conviction without awareness, and a perfectly consistent story that just happens to be wrong.

And just as Leonard lived in a perpetual present , reconstructing his story from scraps , LLMs exist in an eternal “now.” They have no lived sense of before or after, only text patterns suggesting time has passed.

Researchers are now trying to give them longer-term memory: embeddings that persist across sessions, fine-tuning loops that retain experience, and systems that remember what worked last time. But we’re still early. For now, each conversation is another Polaroid — vivid, intelligent, but fleeting.

Until we solve memory, LLMs will remain brilliant amnesiacs: endlessly rediscovering the world they already knew, one prompt at a time.