AuthorsSen Hu, Zhiyu Zhang, Yuxiang Wei et al.
TL;DR
CLONEMEM uses hierarchical life arcs plus evidence-grounded QA to show that flat retrieval beats abstractive memories on long-term AI clone tasks (e.g., Recall-Any-Any 0.6103 vs 0.3913 for A-Mem).
Read our summary here, or open the publisher PDF on the next tab.
THE PROBLEM
Existing memory benchmarks rely exclusively on user–agent conversational histories, which are “temporally fragmented and insufficient for capturing continuous life trajectories.”
This causes AI clones to miss gradual changes in experiences, emotions, and opinions, limiting life-grounded personalization and making long-term trajectory tracking unreliable.
HOW IT WORKS
CLONEMEM centers on Persona and Macro-Level Life Arcs, Meso-Level Phase Generation, and Micro-Level Digital Trace Generation to build coherent long-term narratives.
You can think of CLONEMEM like a layered card catalog for a person’s life: top shelves store life arcs, middle shelves store phases, and bottom drawers store daily traces.
This hierarchical design lets CLONEMEM represent how experiences drive emotional and opinion changes over years, something a plain context window or chat log cannot capture.
DIAGRAM
This diagram shows how CLONEMEM generates personas, life arcs, phases, events, evidence, and final digital traces.
DIAGRAM
This diagram shows how CLONEMEM indexes traces, retrieves with different memory systems, and evaluates answers.
PROCESS
Persona and Macro-Level Life Arcs
CLONEMEM first builds Persona and Macro-Level Life Arcs using Big Five traits and coarse major events to constrain long-term trajectories.
Meso-Level Phase Generation
CLONEMEM refines major events into Meso-Level Phase Generation with persistent internal states that carry evolving emotions and opinions across phases.
Micro-Level Digital Trace Generation
CLONEMEM performs Micro-Level Digital Trace Generation, simulating diaries, social media posts, and messages grounded in detailed events and evidence.
Evaluation Question Construction
CLONEMEM runs Evaluation Question Construction, creating evidence-gated QA instances that test experience, emotion, and opinion trajectories over one to three years.
KEY CONTRIBUTIONS
Long-term individual modeling beyond conversations
CLONEMEM extends memory evaluation from chat histories to non-conversational digital traces using Persona and Macro-Level Life Arcs over 1–3 year trajectories.
Hierarchical life arc data construction
CLONEMEM introduces Meso-Level Phase Generation and Micro-Level Digital Trace Generation to maintain longitudinal coherence in experiences, emotions, and opinions.
Trajectory-focused evaluation tasks
CLONEMEM’s Evaluation Question Construction defines 8 task categories and about 5,000 QA pairs to assess evolving experiences, emotional changes, and opinion dynamics.
RESULTS
Recall-Any-Any k=20 Flat
0.6103
+0.2190 over A-Mem with LLaMA-3.1-8B
Recall-All-Any k=20 Flat
0.2489
+0.0377 over A-Mem with LLaMA-3.1-8B
QA consistency Oracle GPT-4o-mini
0.8325
upper bound vs Flat 0.6955 with GPT-4o-mini
Choice accuracy Oracle LLaMA-3.1-8B
86.06
+8.15 over Mem0 77.91 with LLaMA-3.1-8B k=5
CLONEMEM evaluates AI clones on synthetic bilingual personas with up to 1M-token contexts, focusing on long-term experience, emotion, and opinion trajectories. These results show CLONEMEM exposes a validity–fidelity trade-off where flat retrieval preserves evidence better than abstractive memories, and oracle access to ground-truth context sets a clear ceiling on performance.
BENCHMARK
CLONEMEM evaluates AI clones on synthetic bilingual personas with up to 1M-token contexts, focusing on long-term experience, emotion, and opinion trajectories. These results show CLONEMEM exposes a validity–fidelity trade-off where flat retrieval preserves evidence better than abstractive memories, and oracle access to ground-truth context sets a clear ceiling on performance.
BENCHMARK
Recall-Any-Any for A-Mem, Mem0, and Flat retriever using LLaMA-3.1-8B at k=20.
BENCHMARK
QA consistency for Oracle, A-Mem, Mem0, and Flat with GPT-4o-mini at k=5.
KEY INSIGHT
CLONEMEM shows the Flat retriever reaches Recall-Any-Any 0.6103 at k=20 with LLaMA-3.1-8B, beating A-Mem at 0.5494 and Mem0 at 0.3913.
This is surprising because abstraction-based systems like A-Mem and Mem0 are designed to improve long-term memory, yet CLONEMEM reveals their lossy compression can hurt evidence-level retrieval.
WHY IT MATTERS
CLONEMEM unlocks evaluation of AI clones on continuous life arcs, testing how systems track evolving experiences, emotions, and opinions over years of digital traces.
Builders can now stress-test memory architectures against high-fidelity, longitudinal trajectories, designing AI clones that preserve evidence-level detail instead of relying on generic narrative priors.
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…