Paper finds persistent trace memory adds little benefit in multi-agent LLM inference

A paper finds that splitting long-context work across agents in a three-tier chain-of-agents system cuts peak KV cache to 14.3 MiB per query, versus about 35 MiB for single-pass and RAG baselines.

The persistent trace-recall tier adds 0.368 MiB of peak cache and changes accuracy by +0.015, with a confidence interval that crosses zero. The authors attribute the weak memory gain to single-question benchmarks, where each item carries its own evidence and is scored independently, leaving stored traces little to recall. Elvis Saravia (@omarsar0) shares the arXiv paper on X.