Mnemo Memory — Benchmarks (LongMemEval-S)
Mnemo Memory scores 85.2% (426/500) on the full LongMemEval-S benchmark under the official LongMemEval judge prompts with a gpt-4o-mini judge. A conservative cross-check under a strict score==1.0 exact-match gate gives 80.8% (404/500). US-hosted and model-agnostic. The Search API runs a consistent ~150ms warm (p50) on live production /v1/search.
Per-category results (official LongMemEval judge)
- single-session-user: 95.7%
- single-session-assistant: 87.5%
- knowledge-update: 85.9%
- temporal-reasoning: 86.5%
- multi-session: 77.4% — current bottleneck
- single-session-preference: 83.3%
- Overall: 85.2%
Two latencies, not to be conflated
SEARCH (the /v1/search product) is a consistent ~150ms warm p50. ANSWER generation — optional LLM synthesis on top — is 3.6s median in fast mode. The retrieval path embeds the query with text-embedding-3-small, then runs hybrid search in parallel (pgvector HNSW vector, Postgres full-text keyword, and fact-aware signals) fused with Reciprocal Rank Fusion, returning ranked memories with citations.
Retrieval recall
Recall@5 is 99.2% (496/500) on LongMemEval-S with a top-50 candidate pool and no LLM reranker. At five chunks recall is already 98.0%; widening the pool tenfold to fifty chunks moves it by 1.2 points.