Tweet by SullyOmarr
June 16, 2025
i think we got "ai memory" all wrong everyone is focused on making models remember facts about you when its way more useful in a work setting ex: when you explain something to a coworker it's basically: - the task (this is what we do now with prompts, memory has very little importance) - how to do the task -> this is where "raw intelligence" matters (reasoning etc) and ability to follow instructions - where to get the required information (context) - this is the important and what everyone is missing about memory e.g (look at notion for HR, and then check slack and linear for xyz) with any agent right now it can do points 1 and 2 pretty well, but fails at 3. you have to explain the same thing over and over each time. its not figuring how do the job better over time we need to solve for an agent doing a task and remembering exactly how it did, what it learned, etc eg: once it figures out that revenue data lives in stripe → sheets → mixpanel, it should just know that for every interaction onwards
- Author
- SullyOmarr
- Date
- June 16, 2025
- Canonical URL
- /tweets/sullyomarr-1934630067667374152-bd87ab