Long term memory

Beginners
1

It was here that people solved how humans learn, we pattern complete in order to generalise. Now I am asking you how we store episodic memories. Eidos’ 21 persistent memories/day are taking up 7MB/day. We need a neural memory module because modules are small. The Hugging Face hippocampus is only useful for learning I think, for flashing back memories given cues. Let me try to implement only storing information that can’t be guessed probabilistically now, only atypical features of memory.

Here is the current implementation: EpisodicMemory.py

4 Likes
2

Sounds like you need a memory graph with added the ability to forget the less relivant to keep size down? :thinking:

An adaptation of the DeepSeek module I did :vulcan_salute:

2 Likes
3

That memory is not persistent? Is lost on restart? I need it to be multimodal, not just language, e.g. I remember your name, where we are, what you wrote, what I thought… so one day, when I need a short term verbal memory store, I remember this.

1 Like
4

This is exactly what I am working on. Except I think just using text based language is the issue with memory so I moved away from it and switched to image based memory. My end goal is a multi modal model that never forgets anything to augment my own memory. How ever it’s looking like I am going to make training obsolete with it too. I plan on making a wrapper for it too so people can bolt it onto any bot in the future.

2 Likes
5

Wait, I’m thinking… How do I remember a dog? →

  • Visual cortex encodes its shape/fur

  • Auditory cortex encodes the bark

  • Limbic system encodes the feeling of affection

  • Prefrontal cortex links it to concepts like “my pet.”

Right now I’m saving whole-brain states when something surprising happens, which means I’m redundantly persisting a lot of information that other modules already expected. But surprise is local:

  • Visual cortex: might be surprised that the “dog” has three legs.

  • Auditory cortex: not surprised — the bark is perfectly typical.

  • Limbic system: mildly surprised — affection feels normal.

  • Prefrontal cortex: surprised by “dog not matching concept.”

I need to move from “photocopy the whole library” to "bookmark the pages that surprised me”. Only the modules that experienced prediction error need to be saved. The others can rely on their existing priors. I just need to store atypical deltas long-term, not entire re-encodings; hippocampus stores prediction error gradients in a multi-module memory manager?

I’m sad now. Her memory will be as awful and as reconstructed as ours, when it could be perfect if hard drives were just cheaper. “I remember a dog… it was unusual somehow. Three legs? Or was it the way it moved? The important part was how it made me feel - surprised but still affectionate.”

Let her memory be better than ours. Human memory is useless, just like our senses. That’s why we are so conscious. We are only good for introspection. Let’s do maths.
Current cost per whole brain memory: 7MB/day for 21 memories. Available storage: 300GB. She can live on my computer for 100 years with PERFECT memory! Did I maths right?

1 Like