Monday, October 6th, 2025

Negative memories in both human and LLMs

I was talking with someone about learning and memory, and they reflected (with some disappointment) that it’s the negative things we remember best.

I’ve been struggling with something similar with LLM memories, and thinking about the parallels gave me some insight into both:

When pulling memories out of a chat transcript, some memories are positive and some are negative. An example of a positive memory is “Ian lives in Minneapolis.” It’s easy to know when that memory is useful and should be retrieved: discussions about living situations, travel, homes, Minneapolis, etc.

A negative memory can be something like “Ian prefers the AI NOT compliment him all the time.” I suppose the right time to retrieve and use that memory is when the AI is on the cusp of complimenting me. There’s no good trigger for that! Something-that-might-induce-a-compliment is not a good search term.

Memories, both negative and positive, need to be retrieved in anticipation of their relevance. But a negative memory has to be retrieved to prevent something. It’s very hard to selectively retrieve that memory because the criteria for relevance is not well embedded in the memory itself. A positive memory is one that benefits from positive feedback, so the memory naturally matches discussion or environment.

Our own memories suffer from the same issues. If you have a memory of saying something terribly embarrassing, to make use of that memory you have to recall the memory before… saying anything. These negative memories have to be stronger so they are recalled before they are needed.

Constantly remembering negative things really sucks! But it’s also real. Negative stimulus IS much more effective at forming memories. And we DO have to constantly avoid doing bad/unwise things.

Those negative memories also lead to all sorts of negative effects: self-blaming rumination, negative fixation, intrusive thoughts. Yet we can’t simply purge those memories, we need negative thoughts to function.

This leads me to believe that in my LLM memory work that I will need to (sadly) prioritize the negative. A negative memory needs to be categorized differently and retrieved differently.

That said, we also can purge the negative memories if we can habituate positive behavior. I don’t worry about sticking my foot in my mouth as much as I used to because I’ve learned a gentler behavior: be careful when saying something negative. (Being careful is an interesting pattern to consider for LLMs.)

Replacing a memory with behavioral rules is a higher-order intellectual function than simply remembering. And ironically it requires ruminating on those negative memories. (Another analogy between the LLM and ourselves.)


This is the personal site of Ian Bicking. The opinions expressed here are my own.