Benchmark Radar Day 90: Personal Memory Needs Permission Checks, Not Just Recall Accuracy

Day 90 of Benchmark Radar. The scan fetched 1,231 items, published 535, and marked 242 as recommended.

Author: Koutian Wu; GitHub: ktwu01

Benchmark Radar is live. Open the dashboard, star the repository, or read the paper. The Hugging Face dataset provides the data.

PGMem turns the safety requirements for an agent’s personal memory into executable checks. It includes provenance-typed records, permission-gated assistance, recall thresholds that trigger abstention, and execution-time consent checks, backed by 19 automated tests and a synthetic dataset covering 240 participants and 17,142 task episodes. The checks reach beyond recall accuracy to what the system can trace, decline, and recheck before acting. PGMem: A Reference Implementation of Provenance-Governed Agentic Memory Support with a Synthetic Dataset and Statistical Evaluation.

AmbiGUI tests whether agents operating graphical user interfaces ask for clarification when an instruction is ambiguous instead of acting immediately. It evaluates a decision boundary that ordinary task-completion scores can miss, whether the agent recognizes that the requested action is not sufficiently specified. That separates agents that resolve uncertainty through questions from agents that merely proceed. facebookresearch/AmbiGUI.

Why this matters

A memory system should be judged on whether it can trace evidence, decline unsupported recall, and recheck consent before acting. The same holds at the interface, where an agent that cannot recognize an underspecified instruction will simply act on it.

Scan coverage

This is a keyword-filtered discovery feed, not a representative survey of AI research. The archived briefing used 126 of 535 captured evidence records. These findings summarize its cited releases; discovery dates do not necessarily equal original publication dates.

Data and briefing: 2026-10-24 archived snapshot.