AuthorsZihao Cheng, Weixin Wang, Yu Zhao et al.
TL;DR
LifeBench uses a cognitive partonomic event hierarchy plus dual-agent simulation to build a dense, year-long multi-source memory benchmark where top systems reach only 55.22% accuracy.
Read our summary here, or open the publisher PDF on the next tab.
THE PROBLEM
Existing memory benchmarks mainly cover semantic and episodic recall, leaving habits and skills under-tested despite dense, long-horizon user histories.
When agents ignore non declarative memory, they fail at aligning schedules with routines, suggesting lifestyle changes, or inferring preferences from fragmented phone artifacts.
HOW IT WORKS
LifeBench builds long-horizon user trajectories using Persona Synthesis, Hierarchical Outline Planning, Dual-Agent Daily Activity Simulation, Phone Data Generation, and Question Answering Generation.
Think of LifeBench as a synthetic “lifelog OS” where a planner writes a year-long storyline, and paired agents simulate daily behavior like RAM and reality-checking disk.
This cognitive, partonomic design lets LifeBench stress-test memory systems on dense, multi-source, year-scale data that a plain context window or single chat log cannot capture.
DIAGRAM
This diagram shows how LifeBench’s Subjective Agent and Objective Agent iteratively generate and validate daily activities from atomic events.
DIAGRAM
This diagram shows how LifeBench accelerates persona to phone data synthesis from 58 hours to 8 hours per user year.
PROCESS
Persona Synthesis
LifeBench samples survey priors and PersonaHub attributes, then uses Persona Synthesis to build rich demographics, habits, and social networks for each user.
Hierarchical Outline Planning
LifeBench runs Hierarchical Outline Planning to select up to 10 plots, generate thematic events, and decompose them into atomic events in a partonomic tree.
Dual Agent Daily Activity Simulation
LifeBench’s Dual Agent Daily Activity Simulation uses a Subjective Agent and Objective Agent to turn atomic events into coherent daily activities with constraints.
Phone Data Generation and QA Generation
LifeBench applies Phone Data Generation to create chats, logs, and health records, then uses Question Answering Generation to produce 2,003 grounded questions.
KEY CONTRIBUTIONS
Human Cognition Inspired Synthesis
LifeBench models declarative and non declarative memory via Hierarchical Outline Planning and Dual Agent Daily Activity Simulation, organizing events in a partonomic hierarchy over a full year.
Dense Multi Source Dataset and Benchmark
LifeBench provides 3.66M average context tokens, 5,149 events and 8,046 phone artifacts per user, plus 2,003 questions across IE, MR, TKU, ND, and UA categories.
Scalable Parallel Synthesis Framework
LifeBench introduces parallel thematic event decomposition and time sliced daily simulation, reducing per user year synthesis time from 58 hours to 8 hours using 24 thread parallelization.
RESULTS
Overall accuracy
55.22%
+14.23 over Hindsight
IE accuracy
62.95%
vs Hindsight at 40.67%
MR accuracy
55.11%
vs MemU at 35.18%
ND accuracy
47.32%
vs Hindsight at 50.35% (−3.03)
LifeBench evaluates MemU, Hindsight, and MemOS on 2,003 questions spanning IE, MR, TKU, ND, and UA. The 55.22% overall accuracy for MemOS on LifeBench shows that current memory systems struggle with long horizon multi source reasoning compared to their ~90% scores on LoCoMo and LongMemEval.
BENCHMARK
LifeBench evaluates MemU, Hindsight, and MemOS on 2,003 questions spanning IE, MR, TKU, ND, and UA. The 55.22% overall accuracy for MemOS on LifeBench shows that current memory systems struggle with long horizon multi source reasoning compared to their ~90% scores on LoCoMo and LongMemEval.
BENCHMARK
Overall accuracy on LifeBench across all 2,003 questions.
KEY INSIGHT
On LifeBench, Hindsight drops from around 90% accuracy on LoCoMo and LongMemEval to just 40.99% overall, despite its sophisticated memory networks.
This is surprising because Hindsight was designed for long term memory, yet LifeBench’s dense multi source traces and non declarative questions expose weaknesses in its retrieval and reasoning.
WHY IT MATTERS
LifeBench unlocks a way to stress test memory agents on realistic, year scale, multi source lifelogs that include habits, skills, emotions, and noisy phone artifacts.
Builders can now design and benchmark memory systems that handle 3.66M token histories, infer non declarative patterns, and reason over fragmented traces rather than clean chat logs.
Guilin Zhang, Wei Jiang et al.
· 2026
A-MAC scores candidate memories using Utility, Confidence, Novelty, Recency, and Type Prior combined by a learned linear admission policy with Algorithm 1 A-MAC Memory Admission. On the LoCoMo benchmark, A-MAC achieves F1 0.583 and 2644 ms latency, improving F1 by 0.042 and reducing latency by 1187 ms compared to A-mem.
Robbyant Team, Zelin Gao et al.
arXiv 2026 · 2026
LingBot-World combines a Data Engine, Fundamental World Model, Action-Conditioned World Model, and Post-Training causal adaptation to turn a 28B-parameter video generator into a real-time interactive world simulator. On the VBench benchmark, LingBot-World achieves a dynamic degree of 0.8857 versus 0.7612 for Yume-1.5, while also improving imaging quality to 0.6683.
Manoj Madushanka Perera, Adnan Mahmood et al.
· 2026
AgenticAI-DialogGen chains ChatPreprocessor, KnowledgeExtractor, TopicAnalyzer, KnowledgeGraphBuilder, PersonaGenerator, DuelingChat Agent, ConversationValidator, ConversationRefiner, QAGeneration, and PostProcessing to turn raw multi-session chats into topic-guided, persona-grounded conversations with explicit short- and long-term memories. On the TGC / KG memory QA benchmark, Mistral-7B fine-tuned within AgenticAI-DialogGen achieves 87.36 F1, compared to GPT-4’s 83.77 F1 in a zero-shot setting on the same task.
Answers use this explainer on Memory Papers.
Checking…