The Forgetting Machine: Can AI Learn to Remember?

In my last feature, I quoted MIT’s finding that most companies see little return from generative AI, and that the core problem was learning: AI that “forgets context, doesn’t learn, and can’t evolve.” Readers rightly asked two questions. If AI forgets, how is it useful at all? And is the MIT number even right?

Both questions deserve a straight answer. The short version: the famous 95% figure is weaker than its headlines, but the forgetting it describes is real. Forgetting does not make AI useless. It limits where AI can go. And in 2026, the fixes are finally arriving, though more slowly than the marketing suggests.

What MIT Actually Found

01 · Evidence · MIT NANDA · 80,000 Hours · Wharton · 2025 to 2026

A weak number about a real problem

MIT Project NANDA’s report, “The GenAI Divide: State of AI in Business 2025”, was published in July 2025 and became one of the most quoted statistics in technology. Its own description of its method is modest: 52 interviews, 153 survey responses gathered at industry conferences, and a review of more than 300 public AI initiatives, presented as preliminary findings.

The 95% comes from a funnel. Of the organisations studied, about 60% looked at custom or vendor-built enterprise AI tools, 20% piloted one, and only 5% reached production with a “marked and sustained” impact on productivity or profit within about six months. As Rob Wiblin of 80,000 Hours pointed out in April 2026, that means roughly a quarter of the companies that actually ran a pilot succeeded, against a demanding bar. Critics also noted that the authors work on agent infrastructure, the very fix the report recommends.

Other studies paint a brighter picture. Wharton’s 2025 AI adoption report found that 74% of enterprises measuring returns from generative AI reported positive ROI, and 82% used it weekly. And MIT’s own data showed something its headline buried: employees at more than 90% of the companies surveyed were using personal AI tools for work.

But the diagnosis in the MIT report holds up. Its interviewees described the problem precisely: “Most GenAI systems do not retain feedback, adapt to context, or improve over time.” One said a tool “repeats the same mistakes and requires extensive context input for each session.” Another: “For high-stakes work, I need a system that accumulates knowledge and improves over time.” That is the forgetting problem, and the rest of the evidence confirms it.

The 95% figure is weaker than its headlines. The forgetting it describes is not.

How Is Forgetful AI Useful?

Very, for the right work. Today’s models arrive knowing an enormous amount about the world. What they do not know is you: your customers, your exceptions, last week’s correction. So their usefulness depends on how much of a task depends on memory.

Wharton found the top uses were data analysis, summarising documents and meetings, and editing and writing. Each of those is self-contained: everything the model needs is in front of it. That is why adoption by individuals has been so fast. A person who brings the context with them, in a prompt or a document, gets a brilliant assistant that needs no history.

Forgetting hurts where work is continuous: a customer relationship over months, a project with a history, a process with exceptions, a team that corrects the same mistake every week. There, a forgetful AI is like a gifted temp who arrives every morning with no memory of yesterday. You can get a lot of work out of a temp. You cannot make one your account manager.

A forgetful AI is a gifted temp who arrives every morning with no memory of yesterday.Kintoro

How True Are the Claims?

02 · Evidence · Chroma · CL-Bench · Anthropic · OpenAI · 2025 to 2026

Bigger windows, real memory features, and a gap in between

Claim one: “Our context window holds a million tokens.” True, and now standard. Anthropic made its 1M-token window generally available at standard pricing in March 2026, and every major frontier lab now offers one. But accepting a million tokens is not the same as using them well. In July 2025, Chroma tested 18 leading models and found that “model performance consistently degrades with increasing input length”, a pattern it called context rot. Simple needle-in-a-haystack tests, the ones vendors often quote, overstate how reliable long context is for real reasoning.

Claim two: “Our agents remember and improve.” Partly true, and newly so. In April 2026, Anthropic launched memory for Claude Managed Agents in beta: memories stored as files that agents read and write, shared across agents with set permissions, with every change in an audit log that can be rolled back or redacted. Rakuten reported “97% fewer first-pass errors at 27% lower cost and 34% lower latency”, and Wisedocs said memory “sped verification up 30%”. In May, Anthropic added “dreams”, a research preview that reorganises a memory store overnight, merging duplicates and replacing stale entries. The day before Anthropic’s launch, OpenAI said its new workspace agents “retain memory across runs” and “improve the more a team uses them.” These are real products. The results are customers’ own claims.

Claim three: “AI that learns on the job.” Not yet, in the full sense. In June 2026, researchers including Berkeley’s Matei Zaharia and Joseph Gonzalez published Continual Learning Bench, testing whether frontier systems improve from experience across six domains. Their finding was sobering: “agents frequently overfit to immediate observations or fail to reuse knowledge across instances,” and simple in-context learning outperformed dedicated memory systems. Andrej Karpathy put it bluntly in October 2025: “They don’t have continual learning. You can’t just tell them something and they’ll remember it.” He estimated about a decade to fix agents’ shortcomings.

The window got bigger. The memory got real. The learning is still mostly a promise.

Six Paths to a Memory

As a Kintoro framework, there are six ways the industry is trying to cure forgetting, from the simplest to the hardest.

• Bigger windows. Put more in front of the model. Available now, but costly at scale and prone to context rot.
• Retrieval. Search company documents and feed the relevant parts in at the moment of need. Available now, and good for knowledge. But it retrieves; it does not learn.
• Context engineering. Anthropic’s guidance treats context as a finite “attention budget”: summarise long sessions (compaction), keep structured notes outside the window, and give sub-agents fresh windows. Available now, and the main reason coding agents can work for hours.
• Agent memory stores. Persistent, permissioned, auditable memory that agents read and write across sessions, now shipping from Anthropic, OpenAI and specialists. In beta and early production in 2026.
• Organisational memory. Decision records, context graphs and skills, the enterprise brain from my last feature: memory that belongs to the company, not to a model or an agent. Early.
• Continual learning in the model. Models that update themselves from experience, such as Google’s Titans and Nested Learning research. In the lab, with “catastrophic forgetting” of old knowledge the central obstacle.

The important point for a business is that the first five do not require a scientific breakthrough. They are engineering and governance. Most of the memory a company needs can be built today, around the model, rather than inside it.

Forgetting does not make AI useless. It makes AI a brilliant temp. Memory is what turns it into a colleague.The central statement

Memory Has Its Own Risks

An AI that remembers can also remember wrongly, remember too much, or remember the wrong person’s details, as in our scenario. Three rules follow, and the better products already build them in.

• Memory must be visible. People should be able to see what an agent has learned and where it came from. Anthropic shows memory updates as session events and keeps an audit log.
• Memory must be correctable. Wrong memories must be easy to edit, delete or roll back, with a record of who did it.
• Memory must be owned. What an agent learns from your work is your company’s knowledge. It should stay within your walls, be returned if you leave, and never train someone else’s model.

There is a fourth rule, which is the theme of every Kintoro feature. The best memory signal is a human correction. When a person edits, approves or rejects an agent’s work, that is the most valuable thing the agent can learn from. A company that keeps people in the loop is not slowing its AI down. It is teaching it.

How We Build It into piMonk

03 · Kintoro · piMonk · October 2026

Corrections recorded. Memory per app. Shared memory in build.

I should declare an interest. At Kintoro we build piMonk, a suite of twelve business applications for mid-market companies, on one rule: agents do the work, people keep the decisions.

In piMonk, every agent action is recorded as “proposed, approved, edited or rejected”, and every item waiting for approval shows “the agent’s reason: why, not only what.” Those corrections are exactly the signal agents need to learn from. Our responsible AI page says “agents learn from the work as it happens,” and our trust page says what they learn from “trains your agents only.” Today, each app remembers its own work: piRevenue holds the history of accounts, and piCS holds guides, tickets and history.

What is not finished, we say plainly: our architecture page lists “no shared memory yet.” A shared business memory, so that what sales learns reaches service, is in build with no date. Its rules are set in advance: “Yours. Returned on request. Never pooled.”

Five Checkpoints to 2035

This is how I expect AI memory to develop, with each checkpoint marked as measured, forecast by an expert, or forecast by Kintoro.

1. 2026

   Memory ships Measured

   Million-token windows become standard. Anthropic and OpenAI ship persistent memory for enterprise agents, with audit logs and rollback. Research shows long context degrades and true continual learning is not solved.

2. 2027

   Memory gets governed Kintoro forecast

   My forecast: memory controls, such as what agents may remember, who can see it, how it is corrected and where it lives, become standard questions in every enterprise AI purchase.

3. 2028

   Agents remember by default Kintoro forecast

   My forecast: most enterprise agents carry governed memory across sessions, and having to re-explain context to an agent becomes as unacceptable as a call centre that loses your file.

4. 2030

   Learning in the model begins Kintoro forecast

   My forecast: the first production models that update themselves safely from a company’s own experience appear in narrow, well-governed domains.

5. 2035

   Karpathy’s decade Expert forecast

   Karpathy’s October 2025 estimate of about a decade to work through agents’ shortcomings, including continual learning, runs out around here.

The Business Conclusion

Do not wait for AI that learns on the job. Use forgetful AI where it shines today, and build the memory around it yourself.

For a business leader, that becomes five moves over the next twelve months:

• Sort your AI work into self-contained tasks, where today’s AI already pays, and continuous work, where it needs memory first.
• For continuous work, write the context down: customer histories, exceptions, procedures and decisions, in a form agents can read.
• Treat every human correction as a lesson. Record approvals, edits and rejections, and feed them back to your agents.
• Ask every vendor four memory questions: what does the agent remember, can we see it, can we correct it, and who owns it?
• Test long context on your own work before trusting it. A bigger window is not a better memory.

Then measure the one number that tells you whether your AI is learning: how often people have to correct the same mistake twice.

From Temp to Colleague

AI that forgets has already changed how people work, one self-contained task at a time. AI that remembers will change how companies work, because companies are made of continuity: relationships, histories, lessons learned the hard way.

The memory is arriving in pieces, and the learning will take longer. In the meantime, the best teacher an AI can have is a person who corrects it, and a record that keeps the lesson. At Kintoro, that is how we build piMonk: agents do the work, people keep the decisions, and every correction becomes something the company remembers.

It forgets. It is still useful. And the companies that teach it to remember will be the ones that pull ahead.

Brilliant temp today. Trusted colleague tomorrow.