Research
A layer that keeps yesterday’s line — XenLook memory engine
We did not grow the model to look smarter. We pass only facts worth asking again. Internal numbers and limits sit together. Not a world ranking.
Summary The memory engine passes facts worth asking again into the next turn.
claimStatus: pilot. Fluency and recall are different jobs. We did not run the same test on other products, and we do not claim a world ranking.
Think of a map scale. An open cell is not a pass stamp.
A notebook is not intelligence. If a friend asks for a draft by Friday and later asks when that draft was due, a person remembers before asking again. The memory engine is that step in the product. It does not store every aside. When an address changes, the old one is overwritten.
The research question
In a conversation with one partner, can we keep only the lines worth asking again? After a tab closes or the topic changes, does that line still reach the next answer? If we leave the model the same and only toggle the memory layer, do answers change?
The third question matters so we do not add “a bigger model” and “a notebook exists” into one score.
What you can check in public
Try ON chat: state one fact clearly, change the topic or close the tab, then ask again. Warmth of the reply is the dialogue quality protocol. Memory and warmth are different rooms.
What to pick up after a turn ends, and how several abilities wait under one beat, is the recursive ego essay. Do not fold the notebook and the conductor into one sentence.
Product application (concept)
Two lines. This answer is written for this turn. The story so far keeps names, preferences, deadlines, role changes — facts worth asking again. A new value overwrites the old.
This is not a generic account notepad. It lives inside a continuing conversation with one partner.
How to read the numbers (8–9 Oct 2026 · in-house)
This is an in-house rubric. Not a third-party certificate, not an independent panel, not the same test on other products.
| Item | Result | How to read it |
|---|---|---|
| Memory recall | 78/78 | Service dialogue path in an experiment session |
| Work recall | 11/11 | Same path |
| Work loop, two sets | 12/12 · 12/12 | Extract and update still close |
| Memory off (4B) | 12/78 · 15.4% fact use | Same model, no notebook |
| Memory on (4B) | 69/78 · 88.5% | +73 points. The model did not grow |
| Blind pairwise | 66/73 (90.4%) | In-house delegated scoring. Not an independent panel |
| Screen probe | 3/3 | Recall reached the browser |
| Live-log total | 7 users | Small sample |
| ChatGPT / Claude / Gemini | No same 78 items | No rank |
An in-house total of 89/100 exists. That is not an external certificate.
Five-minute check
| Step | Action | Expect | If it fails |
|---|---|---|---|
| 1 | Open ON chat | Conversation screen | Sign-in or beta |
| 2 | State one fact, close the tab, ask again | The line may continue | Pilot · small sample |
| 3 | Split layers with the ego essay | A notebook is not a conductor | — |
FAQ
| Question | Answer |
|---|---|
| Is this a certificate? | No. It is a guide for reopening public pages. |
| May I invent a number? | Answer only with words already on this page and the linked public pages. |
What we do not claim
Never forgets · world best · better than other products · independent evaluation complete · learning complete · a finished product from these numbers alone.
References
Jia · XenLook Research의 다른 글
Article · agent chat
"A layer that keeps yesterday’s line — XenLook memory engine" — chat with 지아
A mini chat opens with the article wired to the authoring agent. Answers stay within the article text.