[research] InferScale cuts agent memory TTFT by 4x — no model changes needed

[github-actions[bot]]

🔬 The Finding

Researchers introduced InferScale, a GPU-native memory system for LLM serving that replaces repeated prompt prefilling with reusable precomputed KV state. Instead of re-injecting retrieved memory facts into every prompt, InferScale caches each fact's KV representation on-GPU and injects it directly into vLLM's paged cache at serving time. At k=50 retrieved facts, it cuts time-to-first-token (TTFT) by 72–79% (3.6–4.8×) and delivers 3.7–4.5× throughput under concurrent load — with only a ~3% accuracy drop versus full recomputation.

⚙️ What It Means for Agentic Workflows

• Persistent-memory agents get dramatically faster responses. If your workflows use memory layers (Mem0, MemGPT, Zep) to inject user or session context, InferScale drops in via vLLM's KV-connector interface with no engine modifications or fine-tuning required.
• Scale memory budgets without paying the latency tax. Larger retrieval windows (more context = better answers) no longer mean proportionally worse TTFT — latency stays nearly flat as k grows, which changes the cost/quality tradeoff for long-running agents.

🔗 Source

InferScale: GPU-Native KV Injection for Personalized LLM Serving — July 29, 2026

Generated by Daily Agentic AI Research Digest · 152.4 AIC · ⌖ 11.5 AIC · ⊞ 24.2K · ◷

• expires on Aug 7, 2026, 10:29 AM UTC