Loxley LabsLong-term memory for AI assistants.
A memory system for large language models: what it does today, what it scored against standard retrieval (RAG) and on the public LoCoMo benchmark, and the research directions being worked on. Search, selection and answering run locally on one desktop PC with two consumer graphics cards.
Six capabilities, each with its measurement.
| Capability | Measured |
| Answers questions about the user's life and the people in it | 86.3 out of 100 against 77.0 for standard retrieval (244 questions) |
| Recalls what was decided and when | 78.3 against 71.7 (210 questions) |
| Places events in time | 70.9 LoCoMo temporal questions, against 64.6 (321 questions) |
| Says "I don't remember" instead of inventing | 50 of 50 questions about things never discussed |
| Delivers memory during a conversation | about 9 s from message to memory; the answer starts at once |
| Keeps up with new conversations | 101 conversations added in the first nightly update, 16 October 2026 |
1,000 questions from a real archive.
The question set was sealed by fingerprint before any system answered. 20 AI writers wrote the questions from original passages, 5 reviewers checked each one against its source, 150 were traps, and 21 blind graders marked the answers without knowing which system wrote which.
Overall score, out of 100
Score by category, out of 100
LoCoMo benchmark: 81.0 on all 1,986 questions; 76.2 on the four standard categories.
On all 1,986 questions: 81.0 against 78.9 for standard retrieval (192 better, 154 worse, p = 0.047). On the four standard categories: 76.2 against 73.7 (p = 0.098). Independent runs of published systems under one judge range from 67 to 76; our graders are not the judge those runs used, and about 6.4% of LoCoMo's answer key is wrong, so a gap of under 3 points on this benchmark is within its noise.
LoCoMo score by category, out of 100
About nine seconds from a message to its memory.
| Step | Time | Accuracy |
| Search and rerank | 4.4 s | the answer passage is among the candidates 98% of the time |
| Selection of the passages that answer | about 4.6 s | keeps the answer passage 91 to 94% of the time; 86% without the reranking step |
On an earlier 200-question test the full loop scored 89.2 against 87.5 for standard retrieval.
Brought up to date every night.
The memory is brought up to date every night with the day's new conversations. The first update of this kind, on 16 October 2026, added 101 conversations.
Two places the work is going.
Part of the memory would be learned into the model itself, so the assistant knows a topic directly as well as looking it up.
A small network would join several different models into one system that works as a whole.
What these numbers do not cover.
- The 1,000-question test uses one person's history; other users' histories are untested.
- Questions were written and graded by AI models, with review and audit, not by people.
- The latest on a changed decision (47.3) and answers spread over several conversations (48.0) are the weakest categories.
- LoCoMo used different graders from published work; on its adversarial questions no memory scores 98.7 by abstaining.
- The nightly update currently uses cloud AI models as well as local ones.
- Neither research direction has a result reported here.
Statistics: bootstrap 95% confidence intervals over questions; two-sided sign tests on questions where two conditions received different grades. Figures for other systems are from their authors or from independent runs, as cited in the write-up, and were not re-measured by us.
Further detail is available under a non-disclosure agreement. The write-up above is the public version; contact support to ask.