NTU Introduces Σ-Mem: Online Reliability Memory for Multi-Agent Systems
NanyangTechnologicalUniversity · hf · 2026-07-31
Nanyang Technological University introduced Σ-Mem, an online reliability memory designed to solve the "who to trust" problem in multi-agent systems.
- Core Mechanism: Unlike traditional memory that only preserves interaction content, Σ-Mem records historical competence evidence for individual peers and peer relationship evidence, updating online based on post-decision correctness feedback.
- Mathematical Foundation: It leverages Weyl's inequality to bound the spectral change caused by event-level updates, enabling stable online adaptation without retraining underlying models.
- Applications & Results: It provides a unified interface for residual steering, response-free peer routing, or reliability-weighted voting. Experiments across five Qwen-family models show it outperforms majority voting and generalizes to unseen peers and domains.
More from coding & agent
- Testing Kimi K3 Agent: Automates Game Dev with MCP and Flux — mervenoyann · 2026-07-31
- DeepSeek demonstrates autonomous subagent orchestration without prompts — teortaxesTex · 2026-07-31
- Struggling to find MCP tools? Developer seeks best practices for AI coding agents — kampak212 · 2026-07-31
- Read/write is not enough: A four-dimensional risk model for AI agent permissions — Harshit-24 · 2026-07-31
- Generating synced SFX for videos while preserving voiceover using SFX MCP — Soggy-Weather169 · 2026-07-31
- Open-Source Local AI Desktop Agent 'Skales' Hits 1.2k GitHub Stars — tom_doerr · 2026-07-31