CTWM memory controller cuts LongMemEval tokens 24.48% with accuracy parity

A rank-based memory controller called CTWM cut prompt tokens by 24.48% on LongMemEval with aggregate accuracy parity, according to an arXiv paper (2610.00010) on heavy-tailed memory traces in long-horizon language agents. The same controller reduced tokens by 5.9% on Synthetic Graph World while lowering bottom-half tail prediction error by 13.6%. The paper's stated takeaway is that agent memory retrieval should be audited for core–tail concentration, since semantic policies can overuse a small memory core and silently accumulate errors on rare states, and that allocating context by retrieval rank while retaining summarized tail state can cut cost without sacrificing coverage.

arXiv https://arxiv.org/abs/2610.00010 CTWM memory controller cuts LongMemEval tokens 24.48% with accuracy parity Which summary reads better? Pick one — models revealed after.Both summaries are AI-generated. Mistral Large quota or rate limit — check usage and plan. Original headline: Heavy-Tailed Memory Traces in Long-Horizon Language Agents A rank-based memory controller cut prompt tokens by 24.48% on LongMemEval with aggregate accuracy parity, and by 5.9% on Synthetic Graph World while reducing bottom-half tail prediction error by 13.6%. The practical takeaway is that agent memory retrieval should be audited for core–tail concentration, because semantic policies can overuse a small memory core and silently accumulate errors on rare states; allocating context by retrieval rank while retaining summarized tail state can reduce cost without sacrificing coverage.