Skip to content

Glossary

Zero data retention

A vendor commitment to discard what you send and what comes back as soon as the answer is delivered, rather than keeping either for any period.

In plain terms

The vendor answers your request and then keeps nothing: not what you sent, not what came back. It is the strongest thing a hosted service can promise about your material short of never receiving it, and it is normally offered on business arrangements rather than being the default. It also removes several features you may currently be relying on.

01

Why it matters

Because it converts a long list of awkward questions into one short answer. What is kept, for how long, who could see it, what would be produced if somebody asked, what happens at the end: all of those collapse if nothing is retained. For organisations that keep meeting the same questions from reviewers, obtaining this once is frequently cheaper than answering them repeatedly.

02

How it works

It covers the material rather than every record of the interaction. A vendor still needs to know that a request happened in order to bill you and to run its service, so counts and timings typically persist while content does not. That is the ordinary shape of it and it is worth having stated rather than assumed, because the distinction is exactly the one a reviewer will ask about.

You lose the features that depended on keeping things, and the list is longer than teams expect. Conversation history, anything that remembers earlier work, saved outputs and the ability for support to look at what actually happened all rest on retention. Support in particular becomes harder in a specific way: nobody can examine the request that went wrong, because it no longer exists.

It is normally an enterprise arrangement rather than a setting, which places it in a commercial conversation. That means it arrives alongside negotiated terms and an administered account, and it means a team on individual subscriptions cannot simply switch it on. Where it matters, it is a reason to move tier rather than a preference to configure.

It is a different commitment from not training on your material, and the two are frequently confused. A vendor can decline to train while still keeping your material for a period, and the questions a reviewer asks about disclosure and retrieval turn on the keeping rather than on the training. Obtaining one does not give you the other.

Getting it in writing is the whole point, because the value is in being able to state it. A configuration that currently discards material is true today; a contractual commitment is what somebody can rely on and what you can put in front of a reviewer next year. That distinction is the reason this term exists as a purchasing concept rather than as a product feature.

What it buys and what it costs

What it buys and what it costsPresenting this as a straightforward improvement is the mistake worth avoiding, because the left column is genuinely powerful and the right column is genuinely inconvenient. An organisation that keeps meeting the same reviewer questions gets enormous value from closing them permanently, and one that relies on conversation history to do its daily work will notice its absence immediately. The support line deserves particular attention when the decision is made. A user reports that something behaved oddly last Tuesday, and there is nothing to look at, because the material was discarded exactly as agreed. That is the arrangement working correctly, and it is worth having said out loud in advance rather than discovered by whoever is trying to diagnose a problem with no evidence available to them.Questions that closeWhat is kept, and for how long?Who could see it?What would be produced ifsomebody asked?Things that stop workingConversation history.Anything that remembers earlierwork.Support examining what wentwrong.This is a trade rather than anupgrade, and the right-handcolumn is where teams aresurprised. The last line is theone that bites, because it isinvisible until the day somebodyneeds it.
Presenting this as a straightforward improvement is the mistake worth avoiding, because the left column is genuinely powerful and the right column is genuinely inconvenient. An organisation that keeps meeting the same reviewer questions gets enormous value from closing them permanently, and one that relies on conversation history to do its daily work will notice its absence immediately. The support line deserves particular attention when the decision is made. A user reports that something behaved oddly last Tuesday, and there is nothing to look at, because the material was discarded exactly as agreed. That is the arrangement working correctly, and it is worth having said out loud in advance rather than discovered by whoever is trying to diagnose a problem with no evidence available to them.
03

Seen in the wild

  • Comparing what an assistant commits to about keeping conversations on its administered tier against its individual one.

    ChatGPT
  • Establishing whether a provider reached through one interface retains requests, since the answer belongs to whoever runs the model.

    OpenRouter
  • Running a model on your own hardware, which achieves the same outcome by never sending anything in the first place.

    Ollama
04

Common misconceptions

People assume

It is the same as not training on our data.

In fact

They are separate commitments and a vendor can offer either alone. Not training says your material will not improve their model; this says it will not be kept at all. Questions about disclosure, retrieval and what would be produced if somebody asked turn on the second rather than the first.

People assume

It means the vendor keeps no record of anything.

In fact

Content is discarded and the fact of the interaction generally is not, because billing and operating the service require knowing that requests happened. Counts and timings typically persist, which is ordinary and is worth having stated explicitly rather than discovered when somebody asks.

People assume

We can switch it on in settings.

In fact

It is usually an enterprise arrangement settled in a contract rather than a preference. A team on individual subscriptions generally cannot enable it, which makes it one of the more concrete reasons to move to an administered tier rather than a box to tick.

05

Telling them apart

Zero data retention vs Retention policy

Zero data retention

The specific commitment: nothing is kept once the answer is delivered.

Retention policy

How long things are kept, which is a period rather than an absence.

One is a duration you read; the other is the case where the duration is nil and it is contractual.

06

Questions

What does it actually cover?
The content you send and the content that comes back. Records that a request happened, with counts and timings, generally persist because billing and running the service depend on them. That distinction is ordinary and worth having stated in writing, since it is precisely what a reviewer will ask about.
What do we give up?
Anything that depended on keeping material: conversation history, features that remember earlier work, saved outputs, and support's ability to examine a request that went wrong. That last one is the surprise, because a problem cannot be investigated after the fact when the evidence was discarded by design.
Is it the same as opting out of training?
No, and the two are frequently confused. A vendor can decline to train on your material while still retaining it for a period. Questions about what would be disclosed or retrieved turn on retention rather than on training, so obtaining one commitment does not give you the other.
Why does it need to be in the contract?
Because the value is in being able to state it to somebody else. A setting that currently discards material is true today; a contractual commitment is what a reviewer can rely on next year and what survives an interface being redesigned. That is why this is a purchasing concept rather than a feature.
07

Key takeaways

  • It collapses a long list of retention questions into one short answer.
  • Content goes; the record that a request happened generally stays.
  • You lose history, memory features and support's ability to investigate.
  • It is a different commitment from not training on your material.
  • In the contract rather than in a setting is the whole point.
09

Tools that use this

  • ChatGPT

    Retention commitments compared across administered and individual tiers.

  • OpenRouter

    Whether the provider behind an interface retains requests.

  • Ollama

    The same outcome by never sending anything at all.

Last checked July 2026

All glossary terms