The memory tax, and other cheap checks that win
TL;DR. The memory tax (2610.34073): on LongMemEval, PersonaMem and MINT-Eval, many dedicated agent-memory systems score below naive retrieval over the raw history while costing more. The paper splits net value into memory benefit (questions memory answers that RAG misses) and memory tax (questions RAG gets that memory loses). The tax comes from rewriting history, which drops or distorts facts; retrieving from raw history alongside memory and reranking both cuts it. Thinking-in-Bets (2610.34654): an agent writes its claims, checks and confidence into a private checklist before reading user feedback, then revises only weak claims. False reversals on patch review fall from 12.2% to 0.02% across five Gemini models, training-free. OAT (2607.12747, first covered 10-26) recirculated: failure attribution learned only from ~100 successful trajectories, no LLM call at inference.
Relation to prior wiki pages
- Agent memory: second paper in two days after Raw-RAG (10-27, plain retrieval over chat history beats agent memory) and the 10-23 MemoryTax evidence. The 10-23 Looking Ahead asked a memory vendor to answer with frontier-class readers; still unanswered.
- Same shape as today's harness-verification result: the cheap independent check beats the elaborate built-in mechanism.