Memory Tax paper: many agent memory systems underperform naive RAG over raw history
_reachsumit · x · 2026-10-28
The arXiv paper "Memory Tax: Dissecting the Net Value of Agent Memory" runs a large-scale evaluation across benchmarks and model backbones and finds many dedicated memory systems score below naive RAG over raw history — a negative net value. It splits memory's net value into memory benefit (reorganized facts, distilled preferences) and memory tax (losses when memory processing removes or obscures evidence), studies how the tax arises from system and evaluation sides, and generalizes the concept to more memory applications. The paper argues the field should shift from standalone performance to net value; code and results will be open-sourced.
Related event: Memory Tax Study Finds Most Agent Memory Systems Underperform Naive RAG(3 posts)→
More from coding & agent
- HKUST paper: policy secretly encodes harness preferences; return = harness ceiling minus mismatch — HKUSTGZ · 2026-10-28
- Addy Osmani shares guide to Claude Code Mods: build custom UI without learning the API — addyosmani · 2026-10-28
- 21K-star open-source course teaches harness engineering with 14 lectures and 8 hands-on projects — techNmak · 2026-10-28
- Grok Bot gets a free AI phone front desk that reads texts, voicemails and drafts replies — elonmusk · 2026-10-28
- Sourcing agent team splits work 80/20: agents automate, humans keep negotiation and final calls — nordic_ash · 2026-10-28
- Try SPARQL locally in five minutes with Oxigraph's serverless Python triple store — JeremyCMorgan · 2026-10-28