HN Showcase
Back to browse
Fixing AI memory blind spot on connected facts with benchmark

Fixing AI memory blind spot on connected facts with benchmark

by SachitRafa·May 10, 2026·7 points·2 comments
Visit ProjectView on HN

AI Analysis

●●●BangerBig BrainSlick

Entity graph retrieval beats Zep Cloud 59% to 28% on LoCoMo-10 benchmarks.

Strengths
  • •Three-layer retrieval finds connected facts that pure vector search misses.
  • •Ebbinghaus-inspired decay model automatically prunes stale memories without manual cleanup.
  • •Public benchmark scripts allow independent verification of recall metrics.
Weaknesses
  • •Graph expansion adds latency compared to simple vector-only retrieval.
  • •Requires maintaining three separate index structures for each memory store.
Category
AI/ML
Target Audience

AI agent developers and RAG pipeline builders

Similar To

Zep · Mem0 · LangChain Memory

Post Description

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 ?

Similar Projects

AI/ML●●Solid

The Memory for OpenClaw

Inspectable memory records beat black-box embeddings for AI agent context persistence.

Big BrainNiche Gem
ngaut
202mo ago