Ok, as per usual, I learned some stuffs reading through this, thanks to both of you big brains, I truly appreciate it.
If I understood correctly, the embedding bit lives on OAI's cloud, not locally, a...
Try qwen/qwen3-embedding-4b on OpenRouter. Much cheaper and very good.
Also, no, inquiring off topic things should not impact its capability or “memory”.
That is not how embedding or memory works. Embedding is basically like a short “location” for a piece of text, and similar text/topics have similar locations.
For each memory item, a location is generated. Then every time you ask something, that question is converted to a location, and similar items are found and checked. These embeddings are never part of the memory, but are instead a way to filter them down to items that actually matter for the given question
It can keep completely different topics without it becoming “diffuse” or other nonsense
Also, no, inquiring off topic things should not impact its capability or “memory”.
That is not how embedding or memory works. Embedding is basically like a short “location” for a piece of text, and similar text/topics have similar locations.
For each memory item, a location is generated. Then every time you ask something, that question is converted to a location, and similar items are found and checked. These embeddings are never part of the memory, but are instead a way to filter them down to items that actually matter for the given question
It can keep completely different topics without it becoming “diffuse” or other nonsense
2❤️1
