Glider

Glider is an open-source, single-node vector database written in Rust and published by omerfeyzioglu. Every acknowledged write is durable in S3-compatible object storage before the server responds, while local RAM and SSD act only as read caches. It serves nearest-neighbor search with metadata filters over an HTTP/JSON API, with a Python client and an MCP memory server for AI agents. The latest release listed is v1.0.0.

What It Is

Glider is a vector database for similarity search over embeddings. Users start the server with Docker, create collections with their own dimension and metric (squared Euclidean, Manhattan or cosine), upsert vectors with optional metadata, and query for nearest neighbors. A built-in browser console at /console lets users browse points and run queries. Optional server-side text embedding can use a local ONNX model or an OpenAI-compatible endpoint such as Ollama or OpenAI, while stored data remains vectors.

How Durability and Search Work

Each write batch becomes one immutable log object created conditionally in the object store before acknowledgment. In idle time the server seals logs into immutable packs and indexes and publishes a new root generation. A restarted server waits out the old writer's lease, fences it at the store and replays the log. Queries route to candidate blocks, read bounded byte ranges from S3 or the SSD cache, and rerank full vectors. The server builds a clustered ANN index on its own once a collection reaches 250,000 rows and rebuilds it as the data grows. Write requests carry request IDs so retries within the retained 128-commit window return the original outcome.

Filters and Agent Memory

Metadata filters support equality, sets, existence, numeric ranges and logical combinations. Filtered search is approximate by default, and exact: true returns exhaustive matches. The Python client includes glider-mcp, an MCP server exposing remember, recall and forget tools so MCP clients such as Claude Code, Claude Desktop and Cursor can keep durable memory backed by Glider.

Measured Results and Limits

The project reports, for 1M SIFT vectors on S3 Standard from a c7g.2xlarge in eu-central-1, recall@10 of 0.998, warm-cache query p95 of 29.8 ms and write p95 of 86.9 ms. A multi-tenant test with 10,000 collections on one server is also documented. Documented limits: one server with a single committer per collection, no replication or read replicas, plain HTTP (TLS must be terminated by a reverse proxy), and collection dimension and metric fixed at creation. Cache loss can lower approximate recall.