The system changed in June, and most advice online predates it
Nearly everything written about ChatGPT and memory describes saved memories: a list you built by telling ChatGPT to remember something, then had to maintain yourself. That is now a legacy mode you have to switch back to deliberately.
What replaced it, OpenAI calls dreaming. Instead of writing down only what you ask it to write down, a background process reads across your past conversations and synthesises what it decides matters. OpenAI is blunt about why it changed: saved memories “tend to go stale over time and eventually become incorrect or irrelevant”, and using them “could feel like talking to someone who took a few notes, but still forgot everything that wasn’t written down”.
For training conversations that helps in theory but the same old inconsistencies keep creeping in. Mention in passing to ChatGPT that a session felt harder than it should have, and it might remember. It might then use that information when helping you analyse a training plan for your next block, or it might not. It’s that inconsistency that makes the plan hard to trust.
OpenAI’s own scorecard
OpenAI published an evaluation alongside the June release, measuring memory against three objectives it set itself.
Two things stand out. The improvement is large and real: the bottom row went from 9.4% to 75.1% in two years. The bottom row is also still the weakest, which matters more for athletes than for most users.
Staying current over time is OpenAI’s term for handling facts that change. Its example is a trip, where ChatGPT should stop thinking you are in Singapore once July has ended. For most people that row governs a handful of facts a year. A training profile is almost entirely different and changes on a daily basis. It is a live thing that needs constant supervision. That’s where Athmex comes in.
These numbers come from OpenAI’s own evaluation. The announcement gives the three scores and no methodology, dataset or sample size, so treat them as the vendor measuring its own system: directionally useful, not an audit.
Why training facts are the awkward case
OpenAI’s documentation reaches for an endurance athlete to explain where the old system broke. Memories, it says, “could also contradict one another, such as ‘I’m training for a marathon’ and ‘I sprained my ankle,’ which made personalization less accurate”.
Both statements were true when you made them. Only one is true now, and which one governs depends on a date neither sentence carries.
Most of what an AI assistant needs to know about you behaves the same way. A threshold pace only means something alongside the date you tested it, so January’s figure is not wrong, it is superseded. A 5K PB stands until it doesn’t, and the day it fell matters as much as the time. Your goal race is a fact about the future that quietly becomes a fact about the past. An injury has a start and usually an end, so “I strained my calf” is permanent as a sentence and temporary as a fact.
One runner, five facts, two of them out of date
Take an illustrative runner in the second half of a half-marathon build. Over eight months they have mentioned, at various points in various conversations:
- In January, a threshold pace of 4:10/km from a lab test.
- In March, a calf strain that cut running back for three weeks.
- In April, a 5K personal best of 18:42.
- In May, a retest putting threshold at 4:02/km.
- In June, an entry for a half marathon on 18 October with a 1:28 target.
Every one of those was true when it was said. Two are now wrong: the January threshold has been replaced, and the calf resolved in April. A third has a date attached that is getting closer.
Now ask what this week’s long run should look like. The answer depends on which threshold the assistant reaches for and whether it still believes the calf is a problem. It will not usually tell you which it used, and nothing in the reply looks different when it picks the January number.
That is the problem Athmex is built for. Instead of leaving the assistant to work out which threshold is current, you keep the facts yourself: 4:02/km, tested in May, with 4:10/km kept as the value it replaced. The race is a dated event with a target time. The calf sits in your constraints note until you clear it, which you do when it stops mattering. When Claude or ChatGPT reads your context, it gets that record rather than a synthesis of eight months of conversation.
Athmex does not watch your training or work these values out for you. You enter them and confirm them, which is the point: the record holds what you put in it, and it keeps holding it until you change it. See how Athmex works.
What you can and cannot check
The memory summary is the page where you read what ChatGPT has concluded about you. It is useful, and it is deliberately partial. OpenAI says it “should capture the most important details” but “will not include everything that ChatGPT remembers based on your chats”, and that some details are withheld when ChatGPT “determines they are less relevant or not appropriate to show in this view”. The sources panel carries a matching caveat: it “may not show every factor or source that shaped a response”.
You can ask ChatGPT what it thinks your threshold is, and OpenAI suggests exactly that. You are asking the system to report on itself, though, and the answer is generated rather than read out of a field you can inspect.
Deletion works the same way. Removing something completely means deleting every source where it appears, “including past chats, archived chats, files, the memory summary”. A superseded number can survive because the conversation that produced it still exists.
That is the trade in a synthesis system. Keeping itself current without your input costs you the ability to audit it.
Two different jobs
Conversational memory and an athlete record are not competing versions of the same feature. One personalises a conversation from what you have said. The other defines what is currently true. They overlap less than they look like they do.
| What is being compared | ChatGPT memory | Athmex athlete context |
|---|---|---|
| Who decides what is kept | ChatGPT keeps “the details it determines are most important”, drawn from conversation without you asking. | You do. Every value is entered or confirmed by you, and nothing is added because an assistant inferred it. |
| Seeing everything held | The memory summary “will not include everything that ChatGPT remembers based on your chats”. | Everything you have entered is a field you can open and read. Thresholds, personal bests, race results and weight also keep their earlier values. |
| When a value is replaced | Memory is revised towards the newer information as time passes. | For thresholds, personal bests and weight, the new value carries the date it applied and the previous one is kept. |
| Future races and goals | Can be remembered from conversation. | Stored as a dated event with a target. |
| Injuries and constraints | Can be remembered from conversation. | Kept as a constraints note you maintain, so it applies until you change it. |
| Which assistants | ChatGPT. | Claude and ChatGPT. ChatGPT needs a paid plan for custom connectors; Claude works on a free account. |
| Activity history | Not an activity store. | Not an activity store either. Workouts stay with your training platform and reach the assistant through its own connector. |
Things that help, whatever else you use
Put stable facts in Custom Instructions. OpenAI draws the line explicitly: “For explicit information or instructions, you can add it to your Custom Instructions. For information shared via conversations, ChatGPT can remember relevant details for you.”
Give numbers a date when you say them. “Threshold 4:05/km, tested 12 March” gives the system something to order by; “my threshold is 4:05” does not.
Say when a value replaces another instead of only stating the new one, because supersession is the case the system finds hardest.
Read the memory summary occasionally. You can correct it in place by highlighting the text you want changed.
Use a Temporary Chat when you want an answer that ignores stored context, which is a quick way to see whether a recommendation depends on something remembered.
None of that is specific to training. The guide to using Claude or ChatGPT as a training coach covers what to put in a first message and which questions are worth asking.
Who needs a record, and who does not
If you ask a training question now and then and you mostly want general guidance, ChatGPT’s memory is probably enough. It will know roughly who you are, and roughly is fine for roughly shaped questions.
It changes if you treat an assistant as a standing part of your training and expect a session recommendation to use the threshold you tested in May rather than the one from January, to know when the race is, and to know the calf stopped being a problem in April. That is a different requirement, and it is not really a memory problem. It is a record-keeping problem, and it is the one Athmex exists to solve.
Athmex holds no activity history. Workouts stay with your training platform and reach your assistant through that platform’s own connector, installed separately. The assistant still does the reasoning. Athmex decides nothing; it just makes sure the facts it reasons from are the ones you meant.