RoMem uses continuous phase rotation to expire stale agent memories, 2-3x better temporal reasoning
Roger_M_Taylor · x · 2026-09-17
A joint team from Edinburgh, LIGHTSPEED, and St Andrews published RoMem (EMNLP 2026 Main), a drop-in temporal knowledge graph module for agentic memory — pluggable in three lines of code.
Key ideas:
- Rotate, don't delete: outdated facts rotate out of alignment in complex vector space at query time; nothing is erased
- A pretrained Semantic Speed Gate scores how fast each relation expires — bornin stays near zero, worksat rotates fast, zero-shot to unseen relations
- Time as one continuous function, not a label: any timestamp works, including dates never seen in training
Results: SOTA on temporal KG completion (72.6 MRR on ICEWS05-15); 2-3x MRR and answer accuracy on temporal reasoning (MultiTQ), dominates LoCoMo, with zero drop in static memory. Paper and code are public.
More from coding & agent
- What is luxury? Omoju Miller asks where human code comprehension fits in the agentic era — omojumiller · 2026-09-17
- LLMs Solved Machine Understanding, but Machine Accountability in Agentic SDLC Is Still Unsolved — omojumiller · 2026-09-17
- awesome-llm-apps: 100+ Open-Source AI Agents, Skills and RAG Apps With 139k GitHub Stars — Saboo_Shubham_ · 2026-09-17
- Open-Source Agent Skill Pinpoints Exactly Where Readers Stop Reading Your Draft — Saboo_Shubham_ · 2026-09-17
- Semantic web's unfinished half: LLMs solved machine understanding, accountability didn't — omojumiller · 2026-09-17
- Microsoft's ProgramDistill: 4,063 tasks expose coding agents' interactive blind spots — MicrosoftResearch · 2026-09-17