uteke

Uteke is a local-first memory engine designed to give AI agents persistent memory without cloud dependencies. It ships as a single Rust binary with no Python runtime, API keys, or external infrastructure required. The project targets MCP-compatible agents such as Claude Code, Cursor, and Hermes, and also provides a CLI and HTTP API for direct use. Core functionality centers on hybrid search: vector similarity via HNSW (usearch) combined with SQLite FTS5 full-text search, merged through Reciprocal Rank Fusion. This allows recall by meaning as well as exact keywords. Embeddings run locally through ONNX using an EmbeddingGemma Q4 768-dimensional model, downloaded once on first run. The README reports a LongMemEval-S recall_any@5 of 98.2% and approximately 45 ms warm query latency, with zero LLM tokens consumed per query. Notable features include Rooms for multi-agent shared memory with author attribution, time-travel queries that recall memories as they existed at a past point in time, a relationship graph with typed edges, smart decay and salience scoring, orphan detection, consolidation of near-duplicate memories, a document engine for wiki-style knowledge bases, batch import, JSONL import/export, and view-only API keys. A server mode and official Docker image are available for team or remote deployments. The repository includes a benchmark harness for LongMemEval-S with committed raw outputs and reproduction instructions, plus a built-in `uteke bench` command for local latency and throughput testing. The project is licensed under Apache 2.0 and is at version 0.19.0 according to the README.