RAW-RAG Open-Sourced: 256-Token Windows + Rerank Baseline Challenges Complex Agentic Memory
JeremyCMorgan · x · 2026-10-28
- Implementation and experimental code for the paper 'Is Complex Agentic Memory Always Needed? Naive RAG Is Stronger Than We Thought' is now on GitHub (MIT).
- Core claim: long-term conversation memory may not need complex agentic memory (mutation, structural organization, learned retrieval) — raw history + evidence retrieval + reader reasoning is often stronger.
- Concrete reproducible recipe: 256-token windows, retrieve 128, cross-encoder rerank to 15.
- Author's challenge: if your graph memory can't beat this baseline on your data, the graph is cost without return; if it can, you have a number to defend. Repo includes comparison methods (EverCore) and reproduction scripts.
More from coding & agent
- SSH MCP deprecated, succeeded by Rust rewrite SSH-Connect — modelcontextprotocol · 2026-10-28
- adis MCP connector queries Czech VAT unreliability and bank accounts — modelcontextprotocol · 2026-10-28
- Open-source Deepthink claims agent-swarm test-time scaling lifts small-model scores from 14% to 65% — Ryoiki-Tokuiten · 2026-10-28
- Harmonic launches Aristotle: AI that formally verifies code, starting with C and Verilog — wellecks · 2026-10-28
- Dev Calls on X to Ban AI Slop Reply Accounts Flooding viral Threads — generativist · 2026-10-28
- Grok Build v1.0.72 Cuts MCP Context Usage Nearly in Half With Multi-Agent Terminal Upgrades — XFreeze · 2026-10-28