Memory Tax paper: most agent memory systems still lose to naive RAG
omarsar0 · x · 2026-10-27
A paper from UIUC, Tsinghua and others finds many dedicated agent memory systems underperform naive RAG over raw history on LongMemEval, PersonaMem, and MINT-Eval, at higher cost.
The authors split memory's net value into memory benefit (questions memory answers that RAG misses) and memory tax (questions RAG solves but memory gets wrong). The tax arises because building memory rewrites history: facts get dropped or distorted during extraction, surviving ones become harder to retrieve, and memory systems are more sensitive to backbone swaps than RAG.
A simple fix helps: retrieve from raw history alongside memory and rerank both together. Practical takeaway—benchmark any memory layer against naive RAG before adopting it. Code and results will be open-sourced.
More from coding & agent
- Querio launches: analytics platform that puts your data stack in a git repo for AI agents — thetripathi58 · 2026-10-28
- Stop conflating "harness": labs build the pilot, companies build the mech — robleclerc · 2026-10-28
- MiRL: error-targeted RL cuts terminal agent misbehaviors 43-57% — _akhaliq · 2026-10-28
- Vjeux's AI agents hit 2.6x speedups on Open Kernel Challenge kernels, then stalled — Vjeux · 2026-10-28
- Anthropic Reveals 12 New Prompting Rules for Claude Opus 5.5 — JeremyCMorgan · 2026-10-28
- TraceTrial-Bench separates agent success from harness reliability: 31.1% of runs mislead — _akhaliq · 2026-10-28