Memory compaction for agents has an information-theoretic floor
pentad.ai
1 thread
A recent paper, Context Compaction Theory (arXiv:2608.01326) puts a floor on how much you can compact an agent's memory. They measured Opus 4.8 against the floor. 4.8 lands on the random-guess line for membership queries. Not very good; but at least it's latent and $-spendy?
I ran the same probes against Wunderblock, which is based on vector-symbolic architecture, and got 84% of the floor with no LLM at all. Bloom filters sit at 55%.
LLMs are overused!
It can be that LLms need to used at the right place rather than anywhere