Alok Vasudev
@alok
How defensible is AI memory? If I use ChatGPT and it updates its memory, then I throw the raw chat history into Claude -- how hard is it for Claude to reach "memory parity"? (Assume there's easy import/export of chat history)
7 replies
3 recasts
20 reactions
Tarun Chitra
@pinged
Not really defensible IMO, context pollution is real (e.g. context pollution leading to being unable to get the right answer without resetting to 0)
2 replies
0 recast
3 reactions
YB
@yb
i really like "context pollution" been having to clean up my memory the last few weeks bc i notices it over indexes on random niche points from a single thread and keeps referring back to it even when not relevant
1 reply
0 recast
3 reactions
shoni.eth
@alexpaden
It’s always going to come back to some form of precision accurate context, long context dumps were in some sense a fad, and the memory randomness is an accuracy issue. Context pollution actually originates in adversarial prompting but this is a great use of the term in a broader sense Tho I mostly disagree with the comment beyond it because precision accuracy is in fact quite difficult (proof is in your gpt issues)
1 reply
0 recast
2 reactions
shoni.eth
@alexpaden
Resetting to zero when it retains history of learnings from the conversation (even human prompted) is just a manual solution to the same problem. Works now but it’s not actually ideal
1 reply
0 recast
1 reaction
Tarun Chitra
@pinged
Yeah I don’t disagree; but the RL can get stuck in a minima that it can’t get out of (like the inequality example I had) and you could imagine that memory means you never fully restart and only do “warm restarts” which might get stuck more frequently; KL clipping from PPO and GRPO isn’t fully sufficient
1 reply
0 recast
0 reaction
shoni.eth
@alexpaden
for sure— and memory results themselves, at least in this gpt case, are essentially all human feedback, which we know is one of the least valuable steps in training. agree the noted approaches aren’t sufficient as a memory retrieval layer. we can probably see most of them in action through the current exploratory memory system, which yb is pointing out issues in—even in a chat context, that holds true without some prioritization of recency or prompt techniques. but i think memory is possible, and ultimately, the goal of memory is not to forget—that won’t be perfect. most of the therapy system is really just an attempt to deprioritize our own memories
1 reply
0 recast
1 reaction
