The paper turns a user's unlabeled history into test questions for an AI agent's memory, then tunes that memory design without hand labels or model updates.
Method: MemSteer writes questions from evidence that links facts across records, then checks them with source audits and blind answer recovery.
Optimization: MemSteer-Opt revises memory store, retrieval and context construction, with no changes to model parameters.
Results: Over Base-BM25, ATM-Bench Question Score rises 13.4 points, and RHELM accuracy rises 4.1 and 6.6 points without and with external data.
Comparison: The authors say this supervision is competitive with official-QA-guided optimization.
Sign in to comment.
TL;DR of the paper, generated automatically.
The paper turns a user's unlabeled history into test questions for an AI agent's memory, then tunes that memory design without hand labels or model updates.