How much should an AI fitness product remember about the user?
I’ve been thinking a lot about the difference between personalization and real adaptation in fitness software.
A profile with age, goal and experience level is useful, but it still treats each workout request almost like a new session.
A more interesting model is longitudinal: remember what the person actually completed, which movements and muscle groups were trained recently, how difficult the exercises felt, what restrictions exist, and how nutrition has looked over the previous days.
Then the next recommendation is based on that history instead of starting from zero.
The hard part is finding the right balance. Too little memory makes the product repetitive and generic. Too much poorly structured data can create noise and false precision.
For people building AI or health/fitness products: which user signals have actually made personalization meaningfully better in your experience, and which turned out to be unnecessary?
Replies
Nutrition could help a lot but I'd be careful not to pretend messy data is priecies.
I think difficulty is underrated. if i keep marking an exercise as brutal, the next plan should probably notice that instead of blindly repeating it.
How much nutrition history is actually useful before it becomes noise? i would imagine the last few days matter more than trying to remember everything.
Promomix
Do you think perceived effort is more useful than completion data? someone can finish a workout but still be completely wiped afterwards.
@margret_rhyme Perceived effort is the higher resolution signal right up until it becomes a control surface, and after that it is the least reliable thing you hold.
Anything self-reported degrades the moment the user works out what it does. Mark an exercise as brutal twice, watch next week get easier, and brutal now means I would like a lighter week rather than what it actually felt like. Completion does not degrade that way, because the user is not filling in a field, they are doing the thing or not doing it.
So the version I would build asks for less and watches more. Whether they finished the last block, how long they rested, whether they came back on the day they said they would. None of that is self-reported, and it recovers most of what an effort rating was trying to tell you.
Same test for the nutrition question further up: a signal is only signal if it can change the next decision. Is there a rule that fires on what someone ate, or is it being stored because it is collectible?
I wonder if the best system would also know when not to personalize. sometimes a users recent behavior is an excepion rather than a new preference.
for me, workout difficulty is a big one because the same plan can feel very different week to week.
Buffup.AI
I would rather have 5 reliable signals than 50 noisy ones.
Bababot
I have used fitness apps where I entered a bunch of information once and then it basically disappeared into a database. the real test is whether that history actually changes what i get recommended next.
I would probably remember preferences separately from facts. Doesnt like running is useful; ran twice last month might not be meaningful forever.
This feels like one of those areas where better memory could matter more than a smarter model. a mediocre model with good context might beat a great model starting from zero every session.