Memory design matters more than model size OdysseyBench evaluated multiple retrieval strategies: • Long-context prompting (8K+ tokens)
• RAG with raw session or utterance context
• RAG with summarization: session or chunk Chunk-level summaries outperformed all, with ~75%

→ View original post on X — @godofprompt