Benchmark
How reliably Litica finds the right memory
We gave Litica a memory of 127 things it had already seen, then asked it 100 questions and checked whether it could find the right one. The questions came from every angle: exact names, the same goal worded a different way, the feeling behind a moment, when something happened, questions that jump across topics, and look-alikes designed to fool it.
Two things matter when you ask for a memory: does the right one come back, and does it come back first. Litica put the right memory in the top five results on every single question, and ranked it first 97% of the time. On the rare miss, it was almost always the second result.
Even the hardest questions held up. The look-alikes built to trip it up, and the narrow questions scoped to a single field, still landed in the top five every time. We hold these to a strict floor in testing: if accuracy ever slips below 95%, the build fails.
| What we asked | Found in top 5 | Ranked first |
|---|---|---|
| Exact names and entities | 100% | 93% |
| The same goal, reworded | 100% | 93% |
| The feeling behind a moment | 100% | 100% |
| When something happened | 100% | 100% |
| Questions that span topics | 100% | 100% |
| Within a single field | 100% | 90% |
| Look-alike memories | 100% | 100% |
| Overall (100 questions) | 100% | 97% |
About these numbers: they come from Litica's own test suite, 100 questions over a 127-item memory across nine kinds of recall. Every score is held to a 95% floor that the build enforces automatically. The full question set and scoring details will ship with an upcoming technical writeup. See how it works on the homepage.