Skip to content
Plein air painting of a river carrying leaves from an orchard toward the sea.

AI can learn from your data without training

By Updated #opinions

Disclosure: I’m the founder of Achiral, which builds an emergent memory system for businesses. This essay argues for that category.

Adjusting model weights is just one technique for compressing model behavior on your data. Memory and storage is another.

Every time your business works with an AI system, it supplies examples, corrections, and outcomes. The system can keep what it learned from them in two places. It can fold them into its weights through training or fine-tuning. Or it can write them down as saved examples, rules, and procedures, and read them back on the next task. Both are lossy encodings of experience, and both change what the system does next.

I’m not claiming any provider is doing this with your data. I have no evidence that any frontier lab uses customer memory to train its general models. The claim is narrower: the mechanism is the same, so a promise about one mechanism should not be read as a promise about the other.

Two places to keep what was learned

Prediction and compression are closely linked. Language Modeling Is Compression makes that case for language models: a model that predicts well is one that has compressed its data well. Weights are the compression we usually talk about. They are not the only one.

A saved correction, a summary of past negotiations, or a tested procedure is also a compressed record of what the business taught the system. It is smaller than the original conversations, and it changes the next answer.

To the person using it, ChatGPT or Claude is the inference system. The model underneath, the router, the application, and the infrastructure can be almost anything. Knowledge kept anywhere in that stack is knowledge the system has learned about you.

What differs between the two is control:

In the weights In a memory store
Form Numbers spread across the whole model Records, summaries, and procedures
Inspect a lesson Hard Read it, if the system exposes it
Remove one lesson Hard Edit or delete it
Move to another model No Yes, if you control the store
Who holds it The model’s owner Whoever runs the store

The underlying encoding may be different, but the business knowledge inside the inference system is the same.

A test for “it learned”

Saving a record doesn’t show that anything was learned. A log is storage. The test is counterfactual: remove the stored lessons and see whether behavior reverts. If the agent goes back to repeating the mistakes you already corrected, the memory was carrying what the business taught it.

The same test works for weights. Both count as learning only when experience causes a lasting change in what the system does.

What follows for contracts

If both are ways of keeping what was learned from your data, then “no training on our data” is ambiguous. It may mean no weight updates. It may mean no use of your data to fit or adapt any persistent part of the system. Those are different promises, and the second is the one a buyer usually thinks they were given.

So read agreements by function, not by mechanism. Ask what is retained, in what form, who can reuse it, and whether it covers the lessons derived from your conversations as well as the conversations themselves. Here is what two providers commit to today.

Anthropic’s API memory tool uses persistent files, with the application controlling the storage. Its Commercial Terms, Section B, effective June 17, 2025, make this commitment to business customers:

Anthropic may not train models on Customer Content from Services.

“May not” is a prohibition. A welcome bit of certainty in a probabilistic business. After twenty years of privacy promises, I’d still like to read the fine print.

OpenAI takes a different approach in ChatGPT’s memory synthesis: a background process that revisits chat history and updates the memory state used in future conversations. For its covered business services, the Services Agreement, Section 4.2, effective January 1, 2026, says:

OpenAI will not use Customer Content to develop or improve the Services, unless Customer explicitly agrees to such use.

These descriptions explain how the products retain memory. The terms set limits for the services they cover. OpenAI’s article describes a ChatGPT rollout, so its implementation details should not be assumed to apply to every business product.

As more of your business data flows into these systems, the memory store grows thicker. The AI can get better at giving you the response you expect.

The system learns and improves over time, if it forms and uses those memories well.

Where should that knowledge live? Who gets to use it, and can you take it with you when you switch models?

Start with what an agent learns while doing your work.

The lesson is part of your business

Imagine an agent helping a company negotiate renewals. Over many conversations and corrections, it learns which exceptions require approval, which concessions are acceptable, and when a relationship matters more than the margin on a particular transaction.

At first, that knowledge is scattered across conversations. Later, it may become a summary, an inferred preference, a tested procedure, or a rule that guides the next negotiation.

The procedure might contain no customer names or original conversations. It can still hold valuable knowledge about how the company operates.

The company has spent time teaching the agent, much as it would teach a colleague. The agent should need fewer corrections and handle familiar situations better.

The company should be able to inspect those lessons, revise them, decide where they apply, and keep them when it changes suppliers.

Learning without weights is not new

Keeping experience outside the weights is an old idea. A nearest-neighbor classifier retains labeled examples and uses them to predict the label of a new one, with no weights optimized. The original retrieval-augmented generation research describes a model that combines parametric memory with non-parametric memory, and the two work together to produce the result.

A business buys a system that produces decisions and actions, not a set of weights. Its persistent memory, retrieval process, instructions, and tools all shape those outcomes, so I would assess a “no training” promise at the level of the whole system. If it covers only certain forms of parameter optimization, the provider should say so and address memory formation and reuse separately. Calling the mechanism memory, retrieval, or post-processing doesn’t settle the contractual question.

A correction becomes a remembered rule. A successful sequence becomes a procedure. A failed approach becomes something the agent knows to reconsider. The weights can stay unchanged while the system acquires useful knowledge about your business.

In The components of agent memory, published November 6, 2026, Anthropic describes memory using these phrases:

“knowledge that Claude gains from experience” … “saved as context it can read later”

The same definition excludes changes to model weights. It describes learning that persists through saved context rather than parameter updates.

A business has reason to care about the learned procedure wherever it lives. A pricing rule can be commercially important whether it lives in weights, a file, or a graph. Changing the storage primitive does not erase the business knowledge it carries.

Read the promise at the right level

Return to those promises. Anthropic’s terms cover inputs and outputs and impose confidentiality obligations. OpenAI’s agreement permits specified uses, including providing the service, before stating the restriction quoted above. Both the permission and the restriction matter when evaluating a memory feature.

Neither clause alone proves that an authorized memory feature breaches the agreement. Remembering information to serve one customer can differ from using it to improve a provider’s general models or products. The agreement, feature terms, permissions, and actual use matter.

As a business buyer, I would separate permission to run inference from permission to retain experience and adapt the system. A memory provider should explain what changes, what the agreement counts as training, who benefits, and who authorized the use. The customer can permit specific forms of adaptation, set limits, or keep the memory in its own infrastructure.

An unchanged set of weights doesn’t answer those questions. Calling something learning doesn’t prove it was prohibited, either. The training restriction applies to the activities it covers; separate clauses may govern retention, confidentiality, and reuse. Follow the data from the original record to the derived procedure, its use on another task, and any use for another customer or product.

Before buying a memory product, I would ask what happens to the knowledge it derives from our work.

Does the agreement cover inferred preferences, reusable procedures, and processed traces as clearly as it covers the conversations that produced them? Can those lessons be used only for our organization? What can be retained after we delete a conversation or leave the service?

If the agreement forbids a particular use, a memory feature doesn’t authorize it. Renaming a transcript as a procedure doesn’t grant permission. A breach still needs evidence of both the prohibited use and the contractual obligation.

Set explicit limits on the use of derived business knowledge, then check that the system follows them. A reassuring phrase about training should not have to carry the whole agreement.

Keep memory independent of inference

A business may want one model for analysis, another for coding, and a third for a workflow with specific cost or deployment needs. It should be able to choose without rebuilding everything its agents have learned.

Separating durable memory from inference gives the company a place to govern that knowledge across models and applications. It can decide which lessons to retain, which context to send for a task, and which procedures another agent may use.

I would favor that over making a frontier lab’s memory feature the sole home of a company’s operating knowledge. The company should be able to change models and keep what its agents have learned.

Context sent to an inference provider still falls under its processing and retention terms. Moving the memory store doesn’t make every copy or trace disappear.

The storage documentation is worth reading just as closely. Anthropic’s API memory tool documentation describes:

building up knowledge over time without keeping everything in the context window.

It also states:

Memory lives entirely in your application.

With this tool, the application runs file operations against storage it controls. The developer controls the memory store; the quoted lines don’t establish that Anthropic owns or retains those files. What matters is where the memory lives and who controls its use.

An independent memory vendor deserves the same scrutiny. Independence has value when the customer gets enforceable control, usable exports, clear permissions, and reliable behavior. Achiral has to meet those standards too.

The work beyond a database

A vector database helps find related material. A knowledge graph helps represent relationships. Both can be useful parts of a memory system.

The application still has to make important decisions. Which experience should become a lasting lesson? Which correction replaces an older instruction? When should a rule apply to one customer rather than the whole organization? How do we tell whether remembering it improved the next decision?

Take the negotiation example. Retrieving a past concession is useful. Remembering that it was an exception requiring approval, and preserving that condition when reusing the lesson, takes more work.

The agent also needs to adjust when the company changes its policy. Making old assumptions easy to retrieve can help it confidently repeat an outdated practice.

We are working on making experience useful over time. That means forming lessons, checking where they apply, and revising or forgetting them when the evidence changes.

An advantage we have to earn

Our case to investors rests on doing that work well. The frontier labs’ capabilities give us every reason to show what independent memory is worth.

I think there is room for a company focused on the knowledge a customer accumulates across models, tools, and teams. Good engineering, evaluation, integrations, and an understanding of the work can become a real advantage. We have to show that through fewer repeated errors, fewer corrections, and better responses when the client’s circumstances change.

I would ask buyers to evaluate the memory arrangement separately from the model subscription. Make ownership, permitted use, portability, correction, and deletion part of that decision. Apply the same standard to Achiral as to any frontier lab.

A business should be able to buy better inference whenever it becomes available, while keeping the lessons it has paid to learn.

© 2026 Marvin Danig. All rights reserved.