What it is
Agent memory where the many small decisions leave the language model.
What it does: On the write path, Jev decides what type a new observation is, whether it is redundant, and how it links to stored memories across semantic, temporal, causal and entity graph views. On the read path, it routes the query to graph views, allocates search effort, scores new candidates and decides when the evidence is enough. A separate answer model then writes the reply from the selected evidence.
Why it belongs in the showcase: A full research system, with paper and code, that moves frequent memory control out of the generative loop.
Caveats: Paper-reported numbers on one benchmark; live runs have API cost.
Evidence
Public MIT repo with arXiv paper 2609.23986 (Dongming Jiang, Yi Li, Bingzhe Li, September 2026) and a Hugging Face demo; on LoCoMo with GPT-4o-mini as the answer model, the paper reports an overall score of 0.777 versus 0.700 for MAGMA, a 158 s memory build versus 1,044 s for Nemori, and 0.93 s average query latency versus 1.47 s for MAGMA; Jev decisions use Noul and Choice, and code enforces graph and retrieval limits; a mock mode and local Laya support exist (Laya is not benchmarked in the paper); mem0 wrote it up on X on 2026-10-07.
Caveats
- The benchmark numbers are the paper's own; the README says its example commands do not reproduce the full paper evaluation by themselves
- Live runs send memory text and queries to TypeSafe and may call an answer model; budget the API cost
- Results cover LoCoMo only; the README claims no LongMemEval result
Unofficial case page. Stars / list position observed for heat; do not treat heat as a quality SLA. Review the upstream README before you depend on it.