Open-source Graft fights coding agent amnesia with markdown, claims SWE-bench win over Claude Code
thisdudelikesAI · x · 2026-09-07
An open-source project called Graft tackles a core coding-agent pain point: agents rebuild their understanding of a codebase from scratch every session—grepping, opening files, following imports—throwing away what they mapped an hour ago.
How Graft works:
- Writes the agent's codebase understanding once into linked markdown files in the repo—one node per system, API, or concept, in plain English rather than function-name dumps
- No embeddings, no vector database, no warm index; the agent just opens and follows files
- Turns onboarding into a one-time cost instead of a per-task expense
The author claims Graft beat Claude Code on SWE-bench Verified (real GitHub issues with maintainers' tests)—a single-party claim pending independent verification. Still, the file-based, zero-infrastructure memory approach is an interesting pattern for agent engineering.
More from coding & agent
- Knowledge work is much harder for agents to automate than code, argues Matt Pocock — mattpocockuk · 2026-09-07
- SRMA: A Game-Theoretic Bilevel Approach Brings Convergence Guarantees to Multi-Agent LLM Systems — Yihang Chen · 2026-09-07
- What You Leave Unspecified Is the Agent's Free Variable: Paras Chopra's Framework — paraschopra · 2026-09-07
- Building a compounding AI workflow: bank context and rules so tasks never start from zero — oran_ge · 2026-09-07
- Google AI Search MCP server adds grounded web search via Vertex AI or Gemini — modelcontextprotocol · 2026-09-07
- Asterwise launches Vedic astrology MCP server with real ephemeris calculations — modelcontextprotocol · 2026-09-07