Give coding agents git-native persistent memory
A tool that stores persistent memory for AI coding agents in git so context carries across sessions.
Done with Claude Code
What they did
You build a Go binary, then run a bootstrap command on a project. It creates a knowledge/ folder of Markdown files with YAML frontmatter, an agent skill, and an AGENTS.md. Agents reach the memory through the CLI or an MCP server over stdio, which works with Claude Code, Cursor and Codex. The convention tells agents to search before writing and to load only the index pages they need.
How it went
The README claims search under 300µs, about 4ms to validate 50+ concepts, and an 80% token reduction. Those figures come from the project's own benchmarks and comparison table, and the README doesn't show independent results.
Worth knowing
The repo is new, with one commit and 3 stars. The 80% token claim comes from a benchmark you can rerun with make benchmark against a local LLM.
Try it yourself with Claude Code
Set up persistent memory for coding agents in this repository by storing notes in a [folder name, e.g. .agent-memory] directory tracked in git. Create a simple format for recording decisions, conventions, and open tasks, and add instructions so future sessions read it first and update it at the end. Show me the files you plan to commit before you commit anything.
Source
Read the original