FTA-Mem: Fact-Time-Affect Anchored Memory for Low-Density Long-Term Dialogue
Opens the thread with a memory architecture purpose-built for low-density emotional-support dialogue, not general-purpose recall.
AI Storyline
3 items · 2 sources · 3 days
Operational story trace
Memory Long-Term
Follow in this browser to see new updates on your Live feed.
Latest change
A new arXiv benchmark, UTILMEM, argues existing long-term-memory evals measure only pointwise factual recall and proposes scoring how well an agent actually uses the evidence it retrieves.
Earlier contextThe story so far
Three separate long-term-memory research papers grouped by a shared title phrase rather than one developing story. It opened Aug 17 with FTA-Mem's fact-time-affect anchored memory for low-density emotional-support dialogue, then Agent Zero Memory's Aug 30 case for provenance-aware memory instead of one fixed structure.
Day 1 Monday, Aug 17, 2026
Day 2 Sunday, Aug 30, 2026
Agent Zero Memory: Provenance-Aware Long-Term Memory for LLM Agents
Argues against committing to one memory structure, proposing provenance tracking across fact stores, vector indexes, and other backends.
Day 3 Monday, Aug 31, 2026
UTILMEM: Benchmarking Evidence Utilization in Long-Term Conversational Memory
Reframes evaluation itself: scoring whether an agent actually uses retrieved evidence, not just whether it recalls isolated facts.
What to watch — open questions
- Does FTA-Mem's fact-time-affect anchoring generalize beyond emotional-support dialogue to task-oriented agents?
- Does Agent Zero Memory's provenance tracking add meaningful latency or storage overhead versus a single fixed structure?
- Does UTILMEM report results for today's popular memory frameworks (vector-store or graph-based), or only research prototypes?
How this thread was built
editor wrote TL;DR
Storylines are threaded mechanically from the feed: stories that share a distinctive anchor across multiple days and sources. Each item links to its original source. The evidence trace, current state, and open questions are written by the editor routine and refreshed whenever a new beat lands.