Multi-Agent Memory Inconsistency: Where Factual Conflicts Disappear in LLM Agents
Organizations: UC San Diego · Intel · Hewlett Packard Enterprise
Abstract
Multi-agent delegation lets LLM systems distribute evidence collection across workers with separate contexts. This creates a consistency challenge: facts retrieved by workers may be missing from the evidence available to the lead for a joint conclusion. We introduce a benchmark and evidence-flow profiler for studying multi-agent memory inconsistency through controlled factual conflicts in Harvey LAB document-review tasks. Pooled across three models with comparable generation spending, requiring delegation increases retrieval of both conflicting claims from 50.2% to 68.7% relative to single-agent execution, yet conflict detection falls from 16.2% to 13.2%. Within this setting, workers report both claims for only 35.2% of retrieved pairs, and pooled matched comparisons show higher detection than single-agent execution when one worker reads both claims but lower detection when the claims are split across workers. Worker-to-lead handoffs are a central target for improving conflict detection in multi-agent LLMs.
Figures & tables
| Source Reads | Worker Reports | Lead Evidence | Final Report | Judge | Cost | ||||||||
| Model | Delegation | Total | Same | Split | Total | Same | Split | Total | Same | Split | Both | Pass | ( budget) |
| GPT-5.6 Luna | None | 69.9 | 69.9 | 0.0 | — | 69.9 | 69.9 | 0.0 | 21.2 | 15.8 | 0.96 | ||
| Optional | 67.1 | 65.1 | 2.1 | 0.7 | 0.7 | 0.0 | 63.0 | 63.0 | 0.0 | 26.0 | 23.3 | 0.95 | |
| Required | 78.8 | 32.9 | 45.9 | 24.0 | 18.5 | 5.5 | 24.0 | 18.5 | 5.5 | 11.6 | 10.3 | 1.02 | |
| GPT-6 Luna | None | 59.6 | 59.6 | 0.0 | — | 59.6 | 59.6 | 0.0 | 17.8 | 17.1 | 0.95 | ||
| Optional | 65.1 | 63.7 | 1.4 | 0.7 | 0.7 | 0.0 | 63.7 | 63.7 | 0.0 | 19.9 | 19.2 | 0.95 | |
Multi-Agent Systems
Multi-Agent Systems studies how autonomous agents coordinate, compete, communicate, and learn in shared or interdependent tasks.
See all 726 papers