1 pointby scrobot4 hours ago1 comment
  • scrobot4 hours ago
    Hey HN! I'm the author. I built this while working on a multi-agent incident investigation system — 6 AI agents running in parallel, all needing shared working memory. Existing solutions (Mem0, Zep, Redis Stack) all require REST APIs, external embedding services, or custom SDKs. I wanted something simpler: MEMSAVE key text → +OK, with embeddings computed locally.

    Tech stack: GraalVM native-image, Java 25, ONNX Runtime (all-MiniLM-L6-v2), HNSW via jvector, SIMD-accelerated cosine similarity via Java Vector API. First version on plain JVM was 2x slower than Redis — after native-image + SIMD it jumped to 1.36x faster.

    Detailed write-up with benchmarks: https://scrobot.substack.com/p/agentis-memory-redis-compatib...

    Happy to answer any questions about the architecture or benchmarks.