syncable-dev/memtrace-public

What it solves

Memtrace provides "structural memory" for AI coding agents, preventing them from having to re-read entire files or lose context across sessions. It solves the problem of agents performing "blind refactors" where they break dependencies they cannot see, and eliminates the high cost and latency associated with using LLMs to build codebase knowledge graphs.

How it works

Memtrace converts a codebase into a live, bi-temporal episodic structural knowledge graph. Unlike general memory systems that use LLM inference to extract relationships, Memtrace uses Rust-native Tree-sitter AST parsers to deterministically resolve symbols (functions, classes, interfaces) and their relationships (calls, imports, implementations) locally on the machine.

It adds two critical dimensions to standard AST graphs:

• Temporal Memory: Tracks the full version history of every symbol, allowing agents to query what changed, what is unexpected, or what will break based on six different scoring algorithms.
• Cross-service API Topology: Maps HTTP call graphs between different repositories to detect service dependencies.

Who it’s for

Developers using AI coding agents (such as Claude Code, Cursor, Windsurf, or DeepSeek Harness) who need their agents to have a deep, architectural understanding of large-scale codebases without sacrificing privacy or incurring API costs.

Highlights

• High Performance: Indexes 1,500 files in under 2 seconds with zero API costs, significantly faster than LLM-based memory systems.
• MCP Native: Exposes over 25 tools via the Model Context Protocol for search, relationship analysis, impact analysis, and graph algorithms.
• Broad Language Support: Supports 20+ languages including Rust, Go, TypeScript, Python, and Java, as well as infrastructure-as-code like Terraform and GitHub Actions.
• Local-First Privacy: Runs entirely on the local machine; source code never leaves the environment.
• Agent Skills: Ships with 17 pre-defined skills that teach agents how to chain tools for tasks like incident investigation and refactoring guides.