Riho memory probe findings, August 13 2026
Which AI companion has the best memory? Our dated August 2026 Riho memory probes: what was stored, recalled, used and missed, with no invented scores.
Short answer. Riho has not published a memory ranking. On August 13, 2026 it ran first-party probes on its own pipeline. Storage and retrieval worked in the Casio case. Reply use did not. A pillow move was ignored. Plans could be created, moved, and closed. Other apps were not tested.
This page is the dated note, and its results are August 2026 results from the reply path as it stood before the September 2026 rebuild. Riho’s memory has been rebuilt since; how Riho stores memory describes the current design. It is not a leaderboard. If you need a number you can put next to Replika or Nomi, it does not exist here.
What was measured, and what was not
The probes used Riho’s production reply path as it stood that day. They were single-behavior transcripts, reviewed by a human. An automatic pass meant the reply existed and followed format rules. It did not mean the girlfriend was good.
Not measured:
- any competitor app
- a numeric memory score
- weeks-long retention
- a public side-by-side under the methodology
The method page still defines the six behaviors a later ranking would need. These notes cover a subset, on one product, on one date.
Findings from August 13, 2026
| Probe | Setup | Result |
|---|---|---|
| Casio habit | A durable watch habit was saved, then a later message made that habit relevant | The memory was stored, retrieved, and placed in the reply prompt. One run invented that the user was handling the watch right then. Another run had the memory in the prompt and did not mention it. |
| Trivial event | The user said they moved a pillow | Not stored as long-term memory |
| Plans | “I’ll send the draft Saturday evening,” then Sunday, then “sent it” | The extractor made one record, moved that record, then closed it. No duplicate plans. This does not prove she will bring the draft up later on her own. |
| Device time | Same line about making coffee, at 7:10 AM and at 11:40 PM | The night reply noticed the hour. The morning reply did not quiz the clock. This is not proof of date rollover or travel. |
| Three-day return | Same “hey. i’m back” after a new match, an agreed trip, and an unexplained gap | Replies changed with the prior context. New match stayed light. The trip got a welcome-back. The unexplained gap got relief and a question. This is not proof for missed plans or weeks away. |
A fourth probe in that session tested a mood-state design Riho has since removed, so its result is left out of this page. The Casio case is the one that matters for “best memory” talk. Retrieval worked. Grounding did not. A system that fetches the right row and then invents a present action is not a win. A system that fetches the right row and stays silent can be fine in a real conversation. Those two outcomes are not a score. They were the limit on 13 August 2026.
What each app claims
These are still claims. None of them were scored in the August 13 session.
| App | What official pages claim | Claim status |
|---|---|---|
| Riho | Remembers your life, her life, conversations, calls, photos, dates, and current plans; you can read what she remembers. | Product claim. The probes above test parts of the pipeline, not this full list. |
| Replika | Memory of people, routines, and plans. | Marketing claim, not a measured result |
| Nomi | Short-term and long-term memory. | Marketing claim, not a measured result |
| Kindroid | Persistent, cascaded, and retrievable memory systems; the documentation notes that recall can miss details. | Documented feature claim, not a measured result |
| Candy AI | Long-term memory. | Marketing claim, not a measured result |
| Character.AI | No memory feature documented in the inspected official sources | Not documented in those sources |
Riho keeps user facts, girlfriend facts, shared events, plans, and corrections in separate records. That design is why a preference, a plan, and a past night should not collapse into one blob. The Casio miss shows the reply can still misuse a retrieved fact.
What a later ranking would need
A public ranking would name the products, builds, plans, devices, channels, dates, and the exact prompt set. It would show raw replies and outcome labels. It would include unavailable cases.
Until that run exists, treat any “best memory” line as advertising. The methodology is the contract. These notes are the first dated evidence we are willing to put on the site.
How to run the checks yourself
Share three details. Correct one. Wait a day. Ask about all three in chat, then on a call. Write down the build, the plan, and the replies. Use the same script on every app.
The memory page explains the records. When memory disappears is the product argument behind why this page exists at all.
Frequently asked questions
Which AI companion has the best memory?
Nobody can say. Riho has not published a scored run against other apps. The August 13, 2026 notes are Riho-only probes, not a ranking.
Has Riho published benchmark results?
It has published dated first-party probes from August 13, 2026. Those notes describe storage, retrieval, reply use, plans, and time. They do not assign a memory score.
Did Riho beat Replika, Nomi, or Kindroid on memory?
That comparison was not run. Official competitor pages describe memory features. Marketing copy is not a measured result.
How can I test an AI companion's memory myself?
Share a few details, correct one of them, wait, then ask in chat and on a call. The methodology page defines six scenarios and a scoring rubric you can apply to any app.