Hack Hydra August 2026
A nine-day hackathon challenging developers to build the next generation of AI-powered applications using HydraDB's graph memory infrastructure.
Results
Three tracks, five ranked winners each, plus honorable mentions. The overall top five are drawn from the track winners — each project holds one award, listed both overall and under its track.
Hindsight
Krish Garg
OSS-engine (bitemporal; forked engine)
320/320 AS-OF agreement vs an independent git-lockfile oracle (5,028 snapshots); disclosed a real scaling cliff; patched HydraDB for read_epoch time travel with an upstream PR.
Canon
Team Canon
OSS-engine (OpenCypher/MinIO)
Only T1 entry scored on the OFFICIAL EnterpriseRAG-Bench harness (70.0→82.5% correctness) with a random-filter control landing exactly at baseline (70.0%) proving it cuts the right superseded doc. 511,958 docs indexed; deployed; fail-closed; 69 tests + CI; caught its own judge rate-limit artifact.
substrate-friction
Swyam Sharma
OSS-engine (dual-arm in-engine diff)
Only entry that truly engages the SWE-bench brief: 172 instances, pre-registered studies, negative control (0.545→0.000), retracted-and-kept results, leave-one-repo-out AUC, 4 upstream HydraDB contributions (issues #81/#101/#102, PR #82).
docket
Sergiu
OSS-engine (OpenCypher)
Only entry that ran the FULL 500-question LongMemEval (0.3511 vs 0.000 no-memory baseline); rigorous refusal taxonomy; public per-question inspector; found real HydraDB cursor bug via 207-probe capability map.
blastradius
Arik (Arik Sulistyaningsih)
OSS-engine (algo.MSpaths)
Full 226,833-record OSV dump; all 6 track questions as native traversals; 38 CI runs; scale runs 143/300/600; live OSV cross-check with 2nd independent semver impl.
Recognised on product merit rather than a track slot.
Thred
Nikhil Rajpurohit
Hosted-API (@hydradb/sdk; no OSS engine)
DOES NOT QUALIFY FOR A TRACK: verified in code - hosted HydraDB SaaS only (HYDRA_DB_API_KEY, query{type:memory}/context.ingest), no Bolt/Cypher/nodes/edges, graph built by HydraDB's managed inference. But as a PRODUCT it's excellent: deployed thred.fun, npm-published multi-client MCP server, clean monorepo. Awarded Best Product / Special Mention, not a track slot.
Enterprise knowledge and retrieval — EnterpriseRAG-Bench, Salesforce HERB
Canon
Team Canon
OSS-engine (OpenCypher/MinIO)
Only T1 entry scored on the OFFICIAL EnterpriseRAG-Bench harness (70.0→82.5% correctness) with a random-filter control landing exactly at baseline (70.0%) proving it cuts the right superseded doc. 511,958 docs indexed; deployed; fail-closed; 69 tests + CI; caught its own judge rate-limit artifact.
Redwood Ontology
Lakshay Tuteja & Shaurya Saini
OSS-engine (Bolt; algo.MSpaths/SPpaths)
Genuine scale + hard entity resolution: 511,961 docs, Splink probabilistic linkage with surname-consistency guard that blocked 102,953 bad merges. recall@20 0.766. Docked: answer-accuracy table still pending.
X-Ray Evidence Platform
Jayati Ahuja
OSS-engine (algo.MSpaths + SPpaths)
Only entry that ran actual Salesforce HERB - all 30 products, 5,126-person graph, blinded protocol. Docked: reframes HERB as coordination analytics rather than answering its QA questions.
Glasshouse
Shrujal Ganatra
OSS-engine (self-hosted + Cloud search)
166,429 people / 209,388 name forms; 173 contradictions found before any question is asked; 47% overall + 100% refusal on 100 graded answers; 183 tests; deployed.
hyrev
Yash T
OSS-engine (algo.MSpaths)
Sharpest single idea: structural abstention as edge-existence (293-question test TP=10 FP=0); 1514/1514 deterministic product resolution; graph recall 1.000 vs BM25 0.550; results committed.
Repositories and dependencies — supply chain, code graphs, SWE-bench
Hindsight
Krish Garg
OSS-engine (bitemporal; forked engine)
320/320 AS-OF agreement vs an independent git-lockfile oracle (5,028 snapshots); disclosed a real scaling cliff; patched HydraDB for read_epoch time travel with an upstream PR.
substrate-friction
Swyam Sharma
OSS-engine (dual-arm in-engine diff)
Only entry that truly engages the SWE-bench brief: 172 instances, pre-registered studies, negative control (0.545→0.000), retracted-and-kept results, leave-one-repo-out AUC, 4 upstream HydraDB contributions (issues #81/#101/#102, PR #82).
blastradius
Arik (Arik Sulistyaningsih)
OSS-engine (algo.MSpaths)
Full 226,833-record OSV dump; all 6 track questions as native traversals; 38 CI runs; scale runs 143/300/600; live OSV cross-check with 2nd independent semver impl.
Downstream
Ayush Singh
OSS-engine (algo.SSpaths incoming)
84,163 versions / 332,985 resolved edges; 140 exposed services in 110 ms; found + engineered around the 1024-path cap and the 1M-edge admission refusal; deployed.
PatchDocket
Oliseh Genesis
OSS-engine (algo.SSpaths)
Best-productized code-graph (2B) build: AST→graph, Pareto evals with real Valibot negative-control corpus, MCP + Docker + web review UI.
Blast Radius (Rohit)
Rohit Maruri
OSS-engine
Best HydraDB-vs-SQLite honesty: SQLite wins 2-12x, both engines agree on every count; 381 tests.
blastradius (Pranav)
Pranav Achar
Hosted-API (semantic) + SQLite closure
Flawless independent-oracle eval (p/r 1.0) and live Vercel product; but exact closure runs in SQLite (HydraDB is the semantic layer).
Custodian
Benita
OSS-engine (algo.SSpaths)
Novel mind-virus containment (Anthropic-paper grounded), 122 tests, 2-process cross-HydraDB demo. Excellent + Best-HydraDB-Use candidate, but Track-2 FIT (repos/deps) is a stretch.
Agent memory — LongMemEval, BEAM, MedLoCoMo
docket
Sergiu
OSS-engine (OpenCypher)
Only entry that ran the FULL 500-question LongMemEval (0.3511 vs 0.000 no-memory baseline); rigorous refusal taxonomy; public per-question inspector; found real HydraDB cursor bug via 207-probe capability map.
Fog of War
Kenil Himmatbhai Vagha
OSS-engine (bitemporal)
500 Qs + a genuinely NOVEL measured metric - hindsight look-ahead-bias (42.5%, n=255, Wilson CIs) - with HydraDB commit-time axis load-bearing; 5 bisected upstream bugs.
Errata
Kshitij
OSS-engine (OpenCypher)
Best eval methodology: 150 Qs x 3 seeds, McNemar significance (p=0.049/0.016), validated judge (8.3% false-accept), deployed + MCP, publishes its own failures.
Lethe
Joshua (Dami904)
OSS-engine (algo.SPpaths/SSpaths)
Dual-benchmark: LongMemEval 78-oracle (92%/91% vs 1% baseline) AND BEAM (20 chats); unusually honest (explains an earlier 100% was an artifact).
MedMemGraph
Venkat Sumanth Reddy B
OSS-engine (self-hosted)
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.
Ledger
Shubham (0xshubhs)
OSS-engine
Graph-only memory; 116 oracle @ 48.3%; no vector; honest.
HydraMind
Himanshu Kumar
OSS-engine (Bolt)
Best pure MCP memory PRODUCT genuinely on the OSS engine (8 tools, live Bolt tests, 250-write restart gate); accuracy is only a 6-case smoke.