Repo Memory Server for Coding Agents
SkaKri · reddit · 2026-07-12
The author introduces rag-rat: a repository memory server for coding agents, provided as an MCP server for tools like Codex and Claude Code.
It solves the problem of new sessions always needing to re-understand the "why" of the code, including:
- Key dependencies
- Call hierarchies
- Design constraints, rejected alternatives, and historical decisions
- Repository memory tied to symbols, paths, and commits
Implementation-wise, it maintains a local index containing:
- Source code chunks, symbols, call/type/import graphs
- Git history
- "Memories" that can be returned alongside search results
- Memory drift detection and maintenance
The author also notes:
- Support for Rust, TypeScript/TSX, Kotlin, C/C++, Python
- Compatibility with local embeddings, remote inference endpoints, or purely lexical retrieval
- Read-only access to source code, writing only to its own SQLite
- Great support for multi-worktree collaboration, ideal for running multiple agents simultaneously
The roadmap includes P2P memory sync, enabling memory integration into GitHub CI / Codex review.
More from coding & agent
- Soft Clamp cuts tool-call overuse in multi-teacher distillation, from 13.7% to 9.0% — antgroup · 2026-07-21
- Agent harness memory loss and compaction are still a major usability problem — adityaag · 2026-07-21
- SpecJudge runs locally on Ollama to pick the right-sized AI model for your project — jokiruiz · 2026-07-21
- A developer maps out six design rules for CLIs that humans and AI agents can both use — yujiezha · 2026-07-21
- A coding-agent skill that forces ADHD-friendly, answer-first output — ayghri · 2026-07-21
- A set of agent skills for CAD, robotics, and hardware design — earthtojake · 2026-07-21