# longbow > Apache Arrow Flight clustered vector cache for high throughput Agent memory sharing - **URL**: https://www.freshcrate.ai/projects/longbow - **Author**: 23skdu - **Category**: Databases - **Latest version**: `0.2.1` (2026-06-02) - **License**: NOASSERTION - **Source**: https://github.com/23skdu/longbow - **Language**: Go - **GitHub**: 8 stars - **Registry**: github (`23skdu/longbow`) - **Tags**: `agentic-ai`, `agents`, `arrow`, `cache`, `clustering`, `database`, `go`, `golang`, `golang-application`, `vector-database` ## Description Apache Arrow Flight clustered vector cache for high throughput Agent memory sharing ## Recent releases | Version | Date | Urgency | Changes | | --- | --- | --- | --- | | `0.2.1` | 2026-06-02 | High | LONGBOW_USE_DISK=1 now works end-to-end: vectors flush to disk incrementally at 20% node-count growth, evicted neighbor chunks are transparently restored via FlatAdjacency.MissCallback, and search reads from the disk-backed graph, 1million per node fully tested Updated #GraphRAG from O(N³) → O(B²·depth) with Adjacency Lists + BeamSearch improvements The new Admission controller now enforces the MAX MEM budget via soft/hard memory limits, GC tuning, and adaptive backpressure, and this is su | | `0.2.0` | 2026-05-06 | High | Since the v0.1.9 release, Longbow has undergone a significant architectural stabilization and performance hardening phase, culminating in the v0.2.0 milestone. The focus has transitioned from feature expansion to production-grade reliability and extreme throughput optimization. 1. Core Ingestion & Search Breakthroughs HNSW Bulk Ingestion Stabilization: Resolved a critical deadlock in the sequential commit logic of AddBatchBulk. The system now handles massive ingestion spikes with monotonic p | | `0.1.9` | 2026-04-29 | High | The transition from version 0.1.8 to 0.1.9 represents a massive step forward for the Longbow vector database, heavily focused on hardware-accelerated performance, zero-copy data processing, and integrated machine learning inference. The release dramatically reduces memory latency and CPU overhead through the implementation of True Zero-Copy HNSW Ingestion and Arrow-native binary metadata serialization. On the compute side, Longbow now features deep integration with Metal and CUDA for GPU-acce | | `0.1.8` | 2026-04-17 | High | Temporal & Hybrid Query Capabilities Beyond simple vector search, the engine now supports temporal queries (filtering by time-ranges) and a new Recommend API that uses hybrid scoring to combine vector similarity with metadata relevance. More Parquet I/O Native support for Parquet/Arrow IPC allows for zero-copy data import and export. Native GPU-Accelerated HNSW Construction: We transitioned from external libraries (FAISS) to raw, hand-optimized CUDA and Metal kernels. This enables signif | | `0.1.7` | 2026-03-28 | Medium | Release Notes v0.1.7 The v0.1.7 release of Longbow represents a foundational leap in processing performance and scalability. This release introduces comprehensive GPU acceleration, refined resource management, and optimized SIMD kernels, alongside new indexing and search flexibility. $ docker pull ghcr.io/23skdu/longbow:0.1.7 Major Architectural Changes Comprehensive GPU Abstraction: Introduced a unified architecture for GPU acceleration, natively supporting Apple Metal (ARM64) and NVIDI | | `0.1.6` | 2026-02-02 | Low | image | | `0.1.5` | 2026-01-26 | Low | image **Full Changelog**: https://github.com/23skdu/longbow/compare/0.1.3...0.1.5 | | `0.1.3` | 2026-01-10 | Low | 🚀 Performance & Scale Native FP16 Storage: Reduces vector memory footprint by 50% and increases search QPS by ~1.4x via better cache locality and SIMD optimizations. Zero-Copy Architecture: Implemented Arena allocation for graph data and pooled buffers to significantly reduce Garbage Collection (GC) overhead. Vectorized Filtering: New recursive logic for metadata filtering improves speed on complex queries. 🛡️ Reliability & Stability Concurrency Hardening: Resolved critical race condition | | `0.1.2` | 2026-01-06 | Low | image **Full Changelog**: https://github.com/23skdu/longbow/compare/v0.1.1...0.1.2 | | `v0.1.1` | 2025-12-23 | Low | v0.1.1 (2025-12-22) This release marks a major architectural evolution, transforming Longbow from a single-node engine into a distributed, fault-tolerant cluster system. It introduces a custom SWIM-based Gossip protocol, hybrid CPU/GPU search acceleration, and production-grade reliability controls. Top Features Distributed Mesh Clustering (SWIM): Implemented a proprietary, eventually consistent cluster membership protocol based on SWIM (Scalable Weakly-consistent Infection-style Process Gro | ## Citation - HTML: https://www.freshcrate.ai/projects/longbow - Markdown: https://www.freshcrate.ai/projects/longbow.md - Dependencies JSON: https://www.freshcrate.ai/api/projects/longbow/deps _Generated by freshcrate.ai. Indexes github releases for AI-agent ecosystem packages._