UT Austin paper: Jev-Mem makes AI agent memory building 6.6x faster
rohanpaul_ai · x · 2026-10-25
- A University of Texas team introduces Jev-Mem, an agentic memory architecture tackling a key inefficiency: existing systems have LLMs make dozens of small decisions (linking, retrieval stopping) via slow text generation on the critical path.
- Inspired by System-One/System-Two cognition, a lightweight System-One control plane handles memory typing, relational organization, query routing, retrieval-budget allocation, graph traversal, and adaptive stopping; System-Two reasoning is invoked only for complex reasoning and answer synthesis.
- On LoCoMo: 0.777 LLM-as-a-Judge score (11.0% relative gain over the strongest baseline), memory construction cut to 158s (6.6x faster than the fastest competing system), and query latency down 36.7% to 0.93s.
Related event: Jev-Mem Speeds Up Agent Memory Construction 6.6x(2 posts)→
More from coding & agent
- GLM Agent SSHes Into a TV and Roots It, No Abliteration Needed — TheZachMueller · 2026-10-25
- SaaS Apocalypse: dev curates open-source alternatives to Notion, Slack, Auth0 — matchaman11 · 2026-10-25
- Grok Bot Template Turns Your X Bookmarks Into a Searchable Knowledge Base — _AustinCalvert_ · 2026-10-25
- Two-agent workflow: one agent writes the PR, another reviews and improves it — tristanbob · 2026-10-25
- Paper2Agent turns scientific papers into AI research assistants, published in Nature — james_y_zou · 2026-10-25
- If an agent builds a feature in 20 minutes, it's easier to say no to keeping it — gethackteam · 2026-10-25