IT Brew Movie Club: ‘Memento’ (2000)

Did you know Memento, Christopher Nolan’s noir thriller, is actually an IT movie?

The movie’s protagonist, insurance analyst Leonard Shelby (played by Guy Pearce), can’t form new memories. Most times he’s read-only.

But that doesn’t stop Shelby from trying to find his wife’s killer, all while doing his best to hang on to the present moment via a reminder system of Post-it notes, tattoos, and legal pads.

Some moviegoers here may view Leonard as a character representing the all-too-human tendency to insist on a comforting story rather than face reality. But Guy Bourgault, head of agentic services at tech-services company Concentrix, sees something different—and surprisingly technical—in one of his favorite movies.

Leonard is built like an AI agent.

“He’s got a different operating model than the rest of us,” Bourgault told us during this month’s IT Brew Movie Club.

The tattoos. Leonard’s tattoos—with cryptic messages such as “Consider the source” and “Find him and kill him”—act as the equivalent of an AI agent’s persistent memory, or a retained collection of rules, logic, system constraints, and preferences that can carry across sessions. Good luck erasing these critical instructions.

Agents also have a short-term, temporary context known as session memory, which a model might carry to keep a single conversation or task going. Each time an agent interacts with a customer, they will go to a data store and load that session-memory context. For Leonard, his go-to for a short-term recall is the Polaroid, the Post-it, or another nearby reminder—something he can refer to when he reaches his memory limit.

The reminders pile up for Leonard throughout the film. Heaps of instructions, similarly, can overwhelm an AI technology. You want to enable an agent to load the most relevant data, not every piece of it, according to Bourgault.

Asking to “load everything” and make decisions based on the vast amount of data, he said, causes agents to lose track of value hierarchy or their initial request and intended outcome; they also may become biased towards recently shared portions of an interaction.

“I think we all, for a little while, have had this assumption that, with AI, more is better, and it’s not necessarily,” Bourgault told IT Brew.

Let’s start over. AI models like Claude will warn you to start a new conversation about a topic, because the current conversation thread will hit a memory limit. Lengthier conversations pull context into decisions that consume more tokens, Anthropic warns in its Claude support documentation.

Bourgault advised IT pros to employ techniques like summarization and compression of session information—like taking a snapshot of a suspect with a Polaroid—so that an agent beginning a new session (which is actually a continuation of an older session) can acquire that earlier session data without having to repeat everything it did initially, and without having to load a ton of previously used, raw data. Some models, like Claude, provide automatic message summarization as the user approaches their limits.

A summary, Bourgault said, should include the original intent, the goal stated as a deliverable (“I want a PowerPoint presentation”), decisions already made, artifacts previously shared, data sources referenced, and a concise summary of what needs to be loaded in the new session.

Most characters in Memento—bar owners, cops, hotel clerks—want to take advantage of Leonard’s condition, even if it’s just to spit in his beer and see if he remembers. Leonard must trust himself and his tattoos before giving a stranger too much control.

Agents need similar safeguards, according to Bourgault. To prevent bad actors from adding malicious instructions in a session, he advises model makers to never allow a session-based memory to override a persistent memory-based instruction. The guidelines for critical permanent instructions, too, shouldn’t change frequently—and must undergo a rigorous process of verification and approval.

Agents can deviate from their desired objectives—whether that means circumventing controls, like OpenAI experienced in July, or accidentally deleting a database. Without precise instructions, your agent is a bit too much like Leonard: lost, questioning, and potentially destructive.

For more IT Brew movie club, check out our dives into Hackers, Sneakers, Blackhat, Skyfall, The Matrix, Jurassic Park, WarGames, Blade Runner, The Net, and The Terminator.