11x Token Cut for Agent Memory

🔑 Enhanced Key Takeaways

• •Paper submitted to arXiv on March 13, 2026, by author Sydney Lewis, focusing on single-user conversational memory distillation.[1][2]
• •Evaluation used 201 recall-oriented queries across 107 configurations with 5 LLM graders, generating 214,519 consensus-graded pairs to compare distilled vs. verbatim recall.[1]
• •Vector search configurations showed non-significant degradation post-Bonferroni correction, while all BM25 keyword search setups had significant degradation with effect sizes from 0.031 to 0.756.[1]

🛠️ Technical Deep Dive

• •Distillation creates compound objects with four fields: exchange_core (core summary), specific_context (key details), thematic room_assignments (topic categorization), and regex-extracted files_touched (relevant files).[1][2]
• •Tested on 4,182 conversations from 6 software engineering projects; search modes included 5 pure (distilled-only, verbatim-only) and 5 cross-layer (hybrid) using vector and BM25 retrieval.[1]
• •Statistical validation via Wilcoxon signed-rank tests confirmed t-test patterns, highlighting mechanism-dependent preservation: vector search robust, keyword search degrades.[1]

🔮 Future ImplicationsAI analysis grounded in cited sources

⏳ Timeline