← News

ResearchSources: arXiv · 20 Nov 2026

MemMend routes memory answers through selective editing to cut revision calls

The framework's authors report that gating which answers get revised preserves accuracy while sharply reducing editing calls, though some benchmarks show gaps.

In plain words

AI systems that answer questions from stored memories sometimes revise every answer, which costs computing resources and can break correct facts. MemMend is a modular framework that first decides which answers need fixing, then applies small, exact-text corrections. The authors say it keeps the original answer when no valid repair exists.

What was announced

MemMend, introduced in a new arXiv paper, routes potentially incomplete answers to an editor and represents corrections as bounded exact-text replacements. Only edits with resolvable source coordinates and valid edit operations are accepted, and the editor preserves the current answer when no legal repair is proposed. The authors state that accepted edits require no additional semantic-verification call.

On 1,986 LoCoMo development questions, the paper reports 67.17 conversation-macro F1 and a mean Judge score of 88.35, improving over a verified revision chain by 0.82 and 1.24 points while using 62.37% fewer revision calls. Applied to native PropMem, Mem0, and SimpleMem answers, the editing interface raised F1 by 3.79, 6.44, and 0.42 points respectively. The authors also note trade-offs: routing produced a 0.67-point lower Judge score than ungated local review, ungated editing retained higher F1 in one adaptation, and an external PerLTQA dialogue subset revealed a completeness gap.

What it does not show

  • All figures come from the paper's own abstract; no independent evaluation is available.
  • The abstract reports mixed results across benchmarks, including lower Judge scores in one routing comparison and a completeness gap on PerLTQA, so gains are not uniform.

For security and IT teams

Scope
The source is the paper's abstract only; full experimental details are not covered.

Sources

Source-support review completed 20 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 20 Nov 2026

New to the subject? AI from the Ground Up starts from the beginning.