RealCompanion examines memory across months of real conversations

Preprint · revised 8 October

8 October 2026 · revised preprint

Ten companion relationships

Age group not specified

Research / benchmark / Safety & wellbeing

RealCompanion proposes a benchmark built from ten real human–AI companion relationships spanning up to 120 days. The authors highlight the difficulty of recognising when earlier context matters and avoiding unsupported inferences about a person.

Why it matters

Especially relevant to evaluating memory and continuity over time.

Read with care

Preprint, first submitted 1 October and revised 8 October. This note covers the v3 record and abstract, not a full audit of methods, consent or privacy safeguards. A small selected dataset does not establish general product safety or wellbeing effects.

What was checked

Version 3 abstract and submission history checked on 12 October 2026. Checked 2026-10-12.

RealCompanion · arXiv v3 ↗