AI companion research

Memory is the feature people test most often when they compare AI companions. Apps advertise memory in different words, and the differences show up in how a detail returns later, whether a correction sticks, and whether a conversation continues during a call.

Riho publishes its testing method openly so you can run the same checks on any app. This section is where the evidence and results live.

The memory benchmark methodology

The AI companion memory benchmark methodology defines six scenarios: recalling personal details, keeping separate people apart, applying corrections, remembering changed plans, continuing a topic across a call, and recalling the current time and context. Each scenario has a repeatable prompt set and a scoring rubric.

The methodology also documents controls that reduce coaching: prompts that do not repeat the user’s original words, checks that run after a gap, and results that record the app version, plan, and model provider used.

Benchmark results status

The August 13, 2026 probe notes are Riho-only. They report storage, retrieval, reply use, plans, and conflict state. They do not assign a score and they do not rank other apps.

• Method: published and repeatable
• Riho-only probes: dated August 13, 2026
• Multi-app ranking: not run

What you can test today

You do not need Riho or any other app to start testing. The memory guide explains the checks you can run in a few conversations: share a preference, correct it later, ask about it after a gap, and try the same topic in a call. Run the same test on every app you compare, and record what you asked and when you asked it.

All research pages

• Why AI companions default to therapy-speak — the research on therapy-speak, and what we tested to make Riho sound like a person instead of a counselor.
• Riho memory architecture — five context layers, evidence-backed memories, and why every memory carries a receipt.
• How Riho understands time — user-local time as truth, no invented reasons for gaps, and the strict lifecycle for plans and dates.
• Anti-rules: how Riho keeps her personality — why “never do X” beats “be X” for AI companion personality.
• How Riho’s character is designed — prose over numeric traits, anti-rules, annotated tone, and the evolution from archetype to a specific person.
• How Riho models emotion and conflict — a 10-dimension emotion state, trust that must be earned, and a conflict arc with guardrails.
• How Riho is tested — the eval suites, the 30-day relationship arc, and the model-by-model results.