Capassa
← The blog
ArticlePerspektivEconomic insight

From numbers to the whole picture: Why AI must remember the right things, not the most

Marit Wetterhus

Marit Wetterhus

CEO

Marit Wetterhus, CEO of Capassa

We talk a lot about how much information artificial intelligence has access to, how quickly it can analyse data and how good its answers can be. But I think we're entering a more interesting phase. What will matter is not how much information an AI agent has access to, but how well it understands the connection between the information, the decisions we make and what we want to achieve. After all, what good is a correct answer to one question if the answer doesn't take into account what we've already decided?

When Capa remembered something I had forgotten

A while ago I was testing our own AI CFO, Capa. I asked it to assess how two new hires next year would affect the budget. Later I was in a customer meeting, showing how Capa can help with investment decisions. We asked a new question: "We're considering buying a van for NOK 800,000 in March. How will that affect our liquidity?" Capa started analysing the investment, but at the same time it pointed out the two hires, which of course had to be included in the assessment of liquidity going forward. I had actually forgotten that I'd given Capa that information. Capa hadn't forgotten.

The example also shows something else. I had only asked Capa to assess the hires. They hadn't been decided. And that's exactly the distinction that matters: The best thing is for the agent to include what it remembers while also asking about the status, for example whether the hires have been decided or are just an idea we're exploring.

From isolated answers to a complete understanding

A good financial decision can rarely be made in isolation. Investments, hires, operations and the company's goals are connected, and every decision affects the room for manoeuvre for the next one. That's why AI agents must not only retrieve numbers, but understand the connections, remember earlier assessments and use them when new decisions are made. They also have to be able to distinguish between what we're considering, what we've decided and what has actually been carried out. A planned hire can be postponed, an investment can be cancelled, and the assumptions can change.

What Anthropic found

Anthropic recently published The components of agent memory, a research article on how agents organise knowledge, learn from experience and correct mistakes. Two findings fit well with this. The first is that more memory isn't better. In an experiment where two Claude models played Pokémon, the older model wrote 413 memory files and earned one badge, while the newer one wrote nine files and earned four. One noted down everything, the other only what the game didn't show on screen: strategies, plans and rules it had worked out itself. The second is that memory can be wrong from the start or become outdated. A wrong note can be read and trusted in later sessions, which is why Anthropic describes how an agent can check its notes against the original logs. It's the same point: AI must not only remember, it must be able to check that what it remembers is still true.

Who can see what Capa remembers?

That's a question I think every leader should ask. For Capa, the answer is that only you can read what Capa remembers, and no one else can see it. You can correct it by telling Capa, and you can delete the conversations you want.

From more data to a better basis for decisions

This isn't just about technology. It's about how we lead businesses. As leaders we make decisions continuously, and we don't always have a full overview of how they affect each other. Imagine if we had an adviser who remembered earlier assessments, understood the company's goals and made us aware of consequences we might not have thought of ourselves. That's the direction I believe AI agents should be developed in. The goal isn't for AI to remember absolutely everything. The goal is for it to remember what's relevant to the decision, be able to verify the information and understand when the assumptions have changed.

The best AI agents of the future may not be the ones that know the most, but the ones that best understand what matters.

Source: Anthropic, The components of agent memory, 6 November 2026. The Pokémon experiment covers one run per model.

Get new analyses in your inbox

One new post a week, never more. No product ads.

Curious about how Capassa works in practice?

Get in touch