Hebbrix

Hebbrix is a drop-in memory layer for AI agents that replaces similarity-based retrieval with outcome-weighted recall. Instead of returning documents that sound related, it surfaces memories that previously led to good outcomes — and lets dead ends decay. Built for developers running production agents, it integrates by changing a single line of code in your existing OpenAI client, automatically injecting relevant context. At its core, Hebbrix uses a 3-tier cognitive memory (short, medium, long-term) with automatic promotion based on access and importance, plus a 5-layer hybrid search engine that combines semantic vectors, BM25, knowledge graph traversal, importance scoring, and recency boosting. A cross-encoder reranker option further refines results for high-precision workloads. The system also auto-extracts knowledge graphs from natural language without any schema setup, and features a self-improving retrieval system with six quality checks and reinforcement learning after each interaction. Recent v2.4.0 additions include a smart memory ingestion pipeline with a noise-rejection classifier and automatic fact extraction using gpt-5-nano, plus a SEARCH_MIN_SCORE setting to filter weak matches. Hebbrix supports document and media upload (PDF, DOCX, audio, video) up to 100MB, with automatic chunking and transcription. It offers a generous free tier (1,000 credits/month) and scales to enterprise with SOC 2/HIPAA compliance and custom SLAs. Compared to alternatives like vector databases or simple key-value stores, Hebbrix uniquely prioritizes outcome success over similarity, making it a strong choice for long-running agents that need to compound learning. It currently leads the LOCOMO long-term memory benchmark with a 0.0% score — a sign of near-perfect recall on the standard conversational memory test.