Memtrace

What Memtrace does

Memtrace turns a codebase into a live knowledge graph covering every function, class, call edge and version, so agents can query structure instead of re-reading files. It is described as structural, with zero LLM calls, and as bi-temporal, which allows time-travel queries about what changed and when. Indexing is built on Rust and Tree-sitter and runs fully locally, and the README says code never leaves the machine.

The pitch is a fleet of coding agents sharing one view of the same repository, with the same call graph, blast radius and history, so refactors can be replayed with knowledge of what depends on what. Recent releases add LeanCTX Native, with compressed reads on get_source_window, single-call directory maps and a token-savings dashboard. Memtrace also runs as a DeepSeek Harness plugin. The official package is memtrace on npm, and the README warns that other projects with the same name are unrelated.

Key features

• Code knowledge graph of functions, classes, call edges and versions
• Bi-temporal time-travel queries
• Structural indexing with zero LLM calls
• Local-first Rust and Tree-sitter indexer
• MCP interface for coding agents
• LeanCTX Native compressed reads and a token-savings dashboard

When to use it

• Check the blast radius of a function before an agent refactors it
• Let several coding agents share the same repository graph
• Ask how a part of the code changed over time

Who it is for: Developers running AI coding agents on large repositories who want shared, queryable code structure instead of repeated file reads.

How it fits with Hermes Agent

The repository description names Hermes among the supported agents, next to Cursor, Claude Code, Codex, DeepSeek Harness, VS Code and Windsurf. It is a general MCP tool, not a Hermes-specific one.

Note: GitHub could not identify the license (NOASSERTION), so reuse terms need to be checked in the repository.

FAQ