Track 3 · #5
MedMemGraph
Awarded to Venkat Sumanth Reddy B
OSS-engine (self-hosted)
MedLoCoMo 0.783 against full-context 0.757 at 8.3x fewer tokens.
Judges’ citation
MedLoCoMo clinical benchmark: 0.783 vs full-context 0.757 at 8.3x fewer tokens; cross-encoder A/B p=0.00016; different-family judge; found Neo4j-driver Bolt handshake incompatibility. Honestly reports win is not yet significant.
Track 3 · Agent memory — LongMemEval, BEAM, MedLoCoMo