Fetching from the wire…
01
02
03
04
05
06
07
08
Research2026-06-07 · source-backed
Multi-turn serving hurts because the KV cache grows linearly with conversation length, choking GPU memory and bandwidth. Tangram spends memory unevenly across the cache instead of treating all tokens equally, cutting the footprint of long sessions. Source: arXiv Directly useful if you self-host a model behind a chat or agent loop and your sessions run long.
Each link below shares sources, entities, or timing with this story.
arXiv 2607.27090 targets the hidden tax in Mem0, MemGPT, and Zep: retrieved memory facts get re-prefilled on every single request. It precomputes each fact's KV representation, stores it on-GPU with a semantic embedding, and injects it into vLLM's paged cache via the KV-connec...
The diagnosis in this paper is better than the fix, and the fix is very good. Recurrent memory agents fail at long context, but not for the reason most people assume. The bottleneck isn't capture. It's retention. Retention falls below 30% at 896K tokens because every consolida...
Addresses the fundamental privacy dilemma: cloud models need data access but enterprises can't share sensitive information. Splits execution between enterprise-side privacy agents and cloud-side capability agents. Directly relevant to AWS AgentCore and enterprise adoption. arX...
Introduces temporal causal diagnostics to distinguish legitimate task execution from injected manipulation in multi-turn agent interactions, plus context purification to neutralize poisoned content. Directly applicable to anyone building agents that call external tools. arXiv...
Alagappan Valliappan shows built-in Multi-Token-Prediction draft heads run full attention over the entire KV cache at every draft step, so at million-token context the "negligibly cheap" draft dominates cost and deep native drafts can go net-negative. Applying a StreamingLLM-s...
arXiv 2607.11683 splits GraphRAG into separate entity-extraction and consolidation stages using DBSCAN-backed deduplication, LLM summarization and Leiden community detection instead of a single pass. Meno-Lite-0.1 (7B) exceeds Qwen2.5-32B by +12.5% relative harmonic mean on KG...
MindPattern daily
One email a day at 7 AM. Sources and a take on every story. Unsubscribe anytime.