- Author
- Bingyu Yan and five co-authors, Beijing Academy of Artificial Intelligence
- Published
- 28 September 2026
- Kind
- arXiv preprint with benchmarks and a new training dataset
- Original
- arxiv.org/abs/2609.34385
- Subject
- File-based agent memory
- Themes
- What lasts
What it is
The authors work at the Beijing Academy of Artificial Intelligence. The paper contrasts Ahead-of-Time memory, built before a request, with Just-In-Time memory. It introduces JAM, the Memory-Gym training data, and a two-stage training method.
What it says
- The authors say memory built ahead of time can discard fine-grained details that later matter.
- JAM has a Memorizer and a Researcher. The Memorizer keeps raw sessions as files in folders with README summaries.
- The Researcher uses open, search and browse tools in a loop. It stops when it judges the evidence sufficient.
- Memory-Gym synthesizes training data with nine task types across six domains.
- The authors report that a 4B-parameter JAM outperforms 7B to 14B trained baselines. This holds across most settings.
- On LongCodeQA, a domain outside the training data, accuracy rose from 54.08% to 75.54%.
What evidence it gives
The paper reports tables on four benchmarks against ten baselines. It adds ablations, timing and a human audit of 200 training instances. LoCoMo overall F1 was 52.09 for JAM against 43.37 for RAG. Code is released.
Interests and limits
The authors built JAM, Memory-Gym and the comparisons, so results are self-reported. The paper is a preprint. JAM adds query-time latency over one-shot systems. The paper says other memory systems stay competitive on dialogue benchmarks such as LoCoMo. The main text was read, but the appendices were only skimmed.
Connections
These links are this wiki's own reading, not claims the sources make.
- Layers of Memory, Layers of Compression. Kellogg treats memory as layers of lossy compression. Yan and colleagues say compressing early discards details that later matter. They keep raw sessions and search them. The choice is about when to pay: when writing, or when asking.
- Hymn to Babylon, missing for a millennium, has been discovered. Yan and colleagues keep raw sessions because later questions may need discarded details. Babylon's hymn returned because many copies survived, and a platform found 30 more manuscripts. Nobody copying them knew the question. Raw copies let a later reader ask it.