Semantix

Semantix is an open source memory kernel for AI agents, written in Go. It captures what happens in coding agent sessions and turns it into reusable, searchable memory so a later session starts with context from earlier ones, rather than repeating the same setup. The kernel extracts typed, scoped semantic slices from session records and stores them in a local database. It supports BM25 and hybrid retrieval and injects retrieved slices in a stable way so provider prompt caching can reuse them, which the project says reduces repeated token spend. Scheduling and speculative read only prefetch are described as early or experimental, and the site states that cost and performance benefits in production still have to be measured on real sessions. Evidence so far comes from repository tests and synthetic demos. It connects to an agent harness through an agent skill, tool hooks or a gateway. It can be used from the command line, with commands to extract slices from a session file and search them, or through a desktop app for macOS and Windows. The desktop builds are unsigned beta releases that need a manual step to open. Semantix is released under the MIT license, with code on GitHub, and is operated by Queshi Intelligence, a Shanghai company. The site is in Chinese and English. No pricing is published, and the software is free to download.