Shared memory that keeps disagreements beats one that picks winners, arXiv study finds
A gated ablation study of multi-agent debate memory reports that passing ranked conflicts to agents outperforms silently resolving them, with caveats about retrieval coverage.
In plain words
When several AI agents debate using a shared memory, systems usually pick one 'winning' fact so agents never see contradictions. The paper's experiments found that instead keeping contradictions visible and ranking them for the agents improved accuracy and reduced damage from poisoned items. The result held across two retrieval setups but comes from one research harness.
What was announced
In the DebateGraph setup described in the arXiv abstract, four backends share one read/write/snapshot contract, letting the authors independently remove five memory mechanisms under a fixed budget, each gated on a pre-declared threshold. Across 39 runs, one mechanism survived: preserving contradictions and passing the ranked conflict set to the debate rather than collapsing it to a winner improved F1 by +0.340 (95% CI [+0.267, +0.414]) and cut attack success from 22 to 7 of 122 poisoned items (McNemar exact p=0.0003) at equal cost, the authors report.
The effect replicated on two retrieval substrates and showed directionally on HotpotQA documents, per the abstract. No other mechanism survived gating, and two performed worse than inert: a write trust-gate and the router's provenance features each cost accuracy (-0.075 and -0.096 F1) once contradictions reached the prompt. The authors attribute the trust-gate failure to structure: poisoned objects are always entities the graph already contains, against 53% of gold, so groundedness scores the lie above a truthful first mention. They also note a harness bottleneck: conflict sets were built only over retrieved edges, and a pre-declared cap left three quarters of planted contradictions unretrievable, first measuring the mechanism at +0.084. The harness, attack set, and logs are released.
What it does not show
- Evidence is limited to the paper's abstract; full methodology and results are not available for verification.
- All figures come from the authors' own harness and runs; no independent confirmation is reported.
- The abstract reports replication only on two retrieval substrates and directional results on HotpotQA documents, not full replication across all settings.
For security and IT teams
- Scope
- Findings apply to multi-agent debate systems with shared memory as tested in the DebateGraph harness, not to agent memory generally.
Sources
Source-support review completed 19 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 19 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.