Memtrace is a structural memory layer for AI coding agents, developed by Syncable and published on GitHub with about 500 stars and active updates in 2026. It turns a codebase into a live knowledge graph that agents can query in milliseconds, instead of re-reading files on every session.
The graph captures every function, class, call edge, and version in the code, and it is bi-temporal, so agents can ask how the code looked at a past point in time as well as how it looks now. Memtrace is MCP-native and works with Cursor, Claude Code, Codex, DeepSeek Harness, Hermes, VS Code, and Windsurf. Notably, it makes zero LLM calls: the index is built structurally with tree-sitter across more than 20 languages, keeping it fast, local-first, and deterministic. The public GitHub repository is the official home for releases and support, and the tool itself is distributed as an npm package written in Rust.
The approach differs from stuffing code into a vector index: because the graph is structural and versioned, an agent can trace exactly which functions call a component before touching it, and can compare the current structure against earlier sessions. Syncable positions this as shared memory for a whole fleet of agents, so knowledge discovered by one agent’s session is available to the next instead of being relearned from scratch.
Memtrace is a commercial product by Syncable, currently in private beta; the GitHub repository and documentation are public, and the tool is distributed under a proprietary license rather than an open-source one.
It is aimed at developers and teams running AI coding agents on large or long-lived codebases, where giving agents durable structural memory prevents blind refactors and repeated rediscovery of how the code fits together.