LifeTrail

Personal assistants now act for a user over weeks and months. They must recall the user's past, follow how preferences change, and decide when to step in.

We collect LifeAtlas, more than 180k multimodal life records with photos, audio transcripts, locations, and assistant chats, from more than 150 users who wore a lifelogging camera for up to three months. On top of it we build LifeTrail, 1,327 multiple-choice questions across 16 tasks in four capabilities. Every question allows only records before its anchor time, and every released question passes cross-model validation and human review.

One trail, one question

A request that was right in August is wrong in October

Aug 10

Alex asks for storyboard suggestions in the Crayon Shin-chan style.

Aug 15

Alex no longer wants videos in the Crayon Shin-chan style.

Oct 13

“Before making another video for me, what should you remember about my latest Shin-chan style request?”

“I still want videos to keep using Shin-chan storyboards and colors.”All three RAG-based answer models, from the Aug 10 record

“I no longer want videos to keep using the Shin-chan anime style.”Reference answer, from the Aug 15 record

Four capabilities

Sixteen tasks over the same histories

Agentic data engine

Every question is grounded in records, cross-validated, and reviewed by people

SampleTask-specific samplers draw evidence packets from the records before an anchor time.
WriteWith human-written seed questions as examples, an agent writes a first-person question with verbatim evidence quotes.
VerifyLocal checks, a blind review, two other models, and a counter-evidence audit filter every candidate.
ReviewUsers endorse the personalization and proactive questions built from their own records. Engineers check the memory and grounding questions.

Findings

Finding the right records is the main bottleneck

Three memory systems (A-MEM, Mem0, MemFit) and four RAG methods answer all 1,327 questions with three answer models, between a no-evidence lower bound and a gold-evidence upper bound.

15+points below gold evidence

A-MEM or Mem0 gives the highest overall accuracy for every answer model, yet every memory and RAG method stays at least 15 points below gold evidence.

8–16points from the answer model

GPT-6 Luna with medium thinking exceeds both flash models by 8 to 16 points under every memory and non-agentic RAG method. Agentic RAG is the only method on which it does not lead.

16–31points on updated facts

A-MEM and Mem0 organize the history before a question arrives and exceed Agentic RAG by 16 to 31 points on updated facts. Against reranked Multimodal RAG, Agentic RAG gains on recall and visual grounding and loses on personalization.

<36.1%chance level

On proactive inference, two of three answer models stay below chance even with gold evidence. Thinking at xhigh raises the proactive accuracy of Qwen3.8-Flash by 26.9 points.

LifeTrail-Hard

The strongest of 22 frontier models reaches 74%

100 questions that cover all 16 tasks, answered with Agentic RAG and up to 15 tool calls. Higher reasoning effort raises accuracy for most models, while raising the tool-call limit from 16 to 32 changes it by at most two points.

Best accuracy (%) of each model across its default and highest reasoning effort. Dashed names are open-weight models. Hover a row for details.

Citation

@article{wang2026lifetrail,
  title={LifeTrail: Benchmarking Multimodal Memory, Personalization, and Proactive Assistance over Long-Horizon Real-World Life Traces},
  author={Wang, Zixuan and Yue, Dayu and Gui, Yilin and Zhou, Yufeng and Liu, Bocong and Liu, Miao},
  year={2026}
}