Show HN: Fixing AI memory blind spot on connected facts with benchmark

  • Posted 2 hours ago by SachitRafa
  • 4 points
https://yourmemoryai.xyz/
Semantic search alone is not enough to capture all connected facts, they will capture the semantically most identical memory only.

Tested on HotpotQA public dataset:

Vector + BM25 + entity graph: BothFound@5 71.5% Vector + BM25 only: BothFound@5 59.5%

Entity graph is the game changer to extract connected facts.

More Benchmark result:

LongMemEval-S: 84.8% recallAll@5 LoCoMo-10: 59% vs zep cloud 28%

What is your approach for connected facts retrieval ?

0 comments