Skip to content

v1.1.0 — 100% on LongMemEval

Latest

Choose a tag to compare

@iikarus iikarus released this 19 Apr 22:22
· 138 commits to master since this release
f03772f

Dragon Brain v1.1.0 — 100% on LongMemEval

100% recall@5 on LongMemEval (ICLR 2025) — the industry-standard benchmark for AI memory systems. 500 questions. 6 categories. No LLM required. Fully local.

Benchmark Results

Metric Score
recall_any@5 100.0%
recall_any@10 100.0%
recall_all@5 99.9%
recall_all@10 100.0%

Dragon Brain is the only system to achieve 100% R@5 without LLM reranking and with a knowledge graph architecture.

Full results: RESULTS.md

What's New in v1.1.0

Retrieval Pipeline

  • 6-channel parallel retrieval — vector + FTS5 + entity + temporal + relational + associative
  • Weighted RRF fusion — per-channel weights set by intent classifier, k=35, PIT normalization
  • Soft intent routing — all channels fire on every query (no hard gate)
  • Cross-encoder reranking — ms-marco-MiniLM on GPU/CPU (optional, not needed for 100%)

New Capabilities

  • FTS5 lexical search — SQLite BM25 catches keyword matches embeddings miss
  • Entity extraction — spaCy NER for graph-native entity-first retrieval
  • SUPERSEDES energy valves — directional energy flow drains stale knowledge
  • Temporal date parsing — automatic date extraction and hard filtering
  • diff_knowledge_state — time-travel diff between two graph snapshots
  • Semantic Radar dashboard — visual discovery of missing relationships
  • Observation-aware embeddings — entity vectors include observation content

Quality

  • 1,281 tests (was 1,166), 0 failures
  • 34 MCP tools (was 33)

Install

pip install dragon-brain
docker compose up -d

Reproduce the Benchmark

python -m benchmarks.longmemeval.runner --dataset oracle