Fetching from the wire…
01
02
03
04
05
06
07
08
Research2026-08-02 · source-backed
arXiv 2607.27919 argues long-term memory should be a separately scalable parametric module rather than entangled with reasoning in one weight set, backed by distributed Faiss indexing and sparse batch-wise loading of kNN distributions. The 410M+6.9B pairing lifts the 17-benchmark average from 29.86 to 37.34, edging past Pythia-12B's 37.24. Adding a 1.7B domain memory to Qwen3 models from 0.6B to 14B gains over 9 points across three domains at every scale. Unlike RAG, the memory is parametric: no retrieval hop at inference.
Each link below shares sources, entities, or timing with this story.
Shared entity: Qwen3 / Same source domain / Shared topic / Earlier coverage
Both cover Qwen3; reported by the same outlet (arxiv.org); overlapping topics (average, benchmark).
Shared entity: Adding / Same source domain / Shared topic / Earlier coverage
Both cover Adding; reported by the same outlet (arxiv.org); overlapping topics (adding, distributed).
Both cover Adding; reported by the same outlet (arxiv.org); overlapping topics (adding, benchmark).
Shared entity: Qwen3 / Same source domain / Earlier coverage / Tension
Both cover Qwen3; reported by the same outlet (arxiv.org); earlier Qwen3 coverage from 2026-04-21.
Shared entity: Adding / Same source domain / Earlier coverage / Tension
Both cover Adding; reported by the same outlet (arxiv.org); earlier Adding coverage from 2026-03-22.
Same source domain / Shared topic / Tension
Reported by the same outlet (arxiv.org); overlapping topics (beat, benchmark, memory); pushes against this story (versus).
Shared entity: Adding / Same source domain / Earlier coverage
Both cover Adding; reported by the same outlet (arxiv.org); earlier Adding coverage from 2026-07-31.
Shared entity: Qwen3 / Same source domain / Earlier coverage
Both cover Qwen3; reported by the same outlet (arxiv.org); earlier Qwen3 coverage from 2026-07-30.