Testing Jev across the memory pipeline's reranking, chunking and recall decisions
Dhravya Shah tested Jev on reranking, chunking, filtering context before extraction and deciding when to recall: it can cut 58% of extraction tokens, but dropping one line of context can turn a memory about the wrong food.
banger research just dropped!! @DhravyaShah tested jev, @typesafeai's new decision model, across the whole memory pipeline: reranking, chunking, filtering context before extraction, and deciding when an agent should recall at all. did you know jev can cut 58% of the tokens going into memory extraction, but dropping one line like "how about indian food?" can turn "i love that stuff" into a memory about mexican food? decision models are insanely fast at yes/no calls and still need the full context to judge meaning. that's why @supermemory leans on learner-1, a small specialized model that makes observation cheap enough to skip the risky cut entirely :)
This post quotes another post: x.com/dhravyashah/status/2103314339239428201