The Memento Technique: does AI remember?

AI is a transformative force for modern companies, but misconceptions about its capabilities persist (even among technology leaders). One of the most common myths is that AI agents “learn” and “remember” like humans. This misunderstanding can lead to unrealistic expectations, flawed strategies, and missed opportunities for real business value.

As technology partners specializing in AI, one of our core responsibilities at CloudX is to help clients separate fact from fiction. I frequently encounter scenarios where assumptions about AI’s out-of-the-box capabilities lead to misjudged timelines and resource requirements to build an agentic solution. For example, a client once requested an AI agent to generate new documents based on thousands of existing files, expecting a rapid, “plug-and-play” solution deliverable in just a few days.

To set the record straight, I’ll try to clarify what truly happens when AI appears to “remember” over time. To illustrate, I’ll draw a parallel with cinema: Memento, Christopher Nolan’s psychological thriller about a protagonist with amnesia.

The short answer: no, AI doesn’t remember

AI agents do not “remember” or “learn” from your data after deployment. Today’s Large Language Models (LLMs) are stateless and read-only. They do not learn from your data or remember past conversations unless specifically engineered to do so. The illusion of memory is just that, not true recollection or awareness.

The Memento Technique: simulating memory in AI agents

Human learning is a dynamic, ongoing process shaped by experience, reasoning, observation, and adaptation to new information. By contrast, what AI agents perform is best described as a simulation of learning.

In an AI agent, the underlying LLM is trained on massive datasets before deployment. During this process, the model is exposed to vast datasets and adjusts billions of internal parameters to improve its ability to predict the next word in a sequence (a process known as statistical pattern recognition). Once deployed, the model does not improve or acquire new knowledge unless it's explicitly retrained or fine-tuned by engineers.

Software developers apply different strategies to create the illusion of memory and learning in AI agents. This is what I like to call “the Memento Technique”.

If you’ve seen Memento, you’ll remember the protagonist Leonard, who suffers from anterograde amnesia (short-term memory loss). To navigate the world and keep important matters present, he relies on tattoos, photographs, and scribbled notes, which help him emulate continuity, even though he has no actual memory.

Current AI agents operate similarly: they do not possess true memory, but compensate by using external systems to mimic memory, learning, and even personality. Some of the techniques involved are:

• Passing relevant information into the agent’s context window (like giving it a snapshot or “tattoo” for each session)
• Using external tools (databases, APIs) to store and retrieve facts
• Engineering workflows that stitch together context for each new interaction.

The illusion is convincing, but it's engineered, not organic.

What is “context” in an AI agent?

Context is the information visible to the model during a single interaction. This “window” is short-lived. Once the session ends, the memory is gone. Everything the agent should “know” in a session must be included in this context.

Also, AI agents can use tools to access documents, databases, or APIs, but developers must explicitly tell the agent what it can access in each session. The agent doesn’t “remember” it has access to these tools from one session to the next unless it’s designed that way.

The art of building AI agents is knowing what information to include, which tools to give to the agent, and how to curate it for each use case. Like in Memento Leonard knew which photo to take or tattoo to get to help him when his memory vanished.

This practice is called context engineering, and it implies carefully selecting and managing the information that creates the emulation of memory.

Why does this matter for AI strategy?

Understanding the limits and capabilities of AI is key for building effective, scalable solutions. When designing AI-powered products, it’s imperative to distinguish between what the technology delivers out of the box and what demands robust engineering and integration. For example:

• If you want an AI agent to recall past decisions, you need to develop a memory layer.
• If you expect the agent to “get smarter” over time, you need explicit retraining or feedback loops.
• If you want consistent behavior across sessions, you need to manage state and context manually.

Understanding these limits is critical for building scalable and reliable AI solutions that deliver measurable business value.

Engineering: the difference between impressive and meaningful AI

Today’s AI agents create a compelling illusion of memory and learning. Given the rapid pace of AI evolution, today’s limitations may change, but these are the current realities. However, it’s important to recognize that the value delivered by AI agents is rooted in sophisticated engineering: a complex system of prompts, custom logic, tools, and external data sources, that enable these advanced behaviors. While the field of continual learning aims to solve these limitations, enabling AI to learn and adapt over time, this is still an area of active research, not a feature in today’s productive LLMs.