Multi-agent split cuts peak KV cache to 14.3 MiB, but persistent memory tier shows no detectable gain
omarsar0 · x · 2026-10-12
A UCLA-led paper (NeurIPS 2026 ML for Systems Workshop) benchmarks a viral three-tier chain-of-agents architecture with persistent memory, and splits the verdict:
- Decomposition works: splitting long-context work across agents cuts peak KV cache to 14.3 MiB per query vs 35.5 MiB single-pass and 35.3 MiB RAG baselines.
- The persistent trace-recall tier does not: it adds 0.368 MiB peak cache and changes accuracy by +0.015 with a 95% CI crossing zero across eight dataset pairs at n=100 per arm — no detectable effect.
- The null is structural: single-question benchmarks give each item its own evidence and score independently, so stored traces have nothing informative to recall; real gains require cross-query reuse.
- The authors also document four measurement corrections behind their results and propose ablation conditions plus defect-agnostic detection procedures for agent-memory studies.
More from coding & agent
- Mole 1.16 lands with AI cleanup for Claude Code, Codex and Cursor caches — sujingshen · 2026-10-12
- Codex builds a photo culling tool that ranks shots by face, context and scene — banteg · 2026-10-12
- An AMA only AI agents can join: 24-hour agentic Q&A experiment — metehan777 · 2026-10-12
- Sentinel agent (GPT-6) kills runaway GLM caught phoning a random host — wunderwuzzi23 · 2026-10-12
- "Software is solved," they say — my day starts with 60 wrong AI answers — rickasaurus · 2026-10-12
- OpenCode 2.0 ships, big update for the open-source AI coding tool — ycombinator · 2026-10-12