CloneMem: Benchmarking Long-Term Memory for AI Clones

TL;DR

CloneMem addresses the challenge of evaluating long-term memory in AI Clones by grounding memory in non-conversational, longitudinal digital traces. It introduces a hierarchical data-construction pipeline that spans macro life arcs, meso-phase states, and micro-level traces to generate trajectory-grounded QA tasks. Through a standardized retrieval-based evaluation, it shows that conventional flat memory often outperforms abstractive memory architectures, revealing a fidelity-utility trade-off in grounding and state tracking. The findings underscore the need for memory systems that preserve trace-level fidelity and explicit internal-state transitions to achieve reliable life-grounded personalization in AI Clones, with CloneMem serving as a benchmark for future work.

Abstract

AI Clones aim to simulate an individual's thoughts and behaviors to enable long-term, personalized interaction, placing stringent demands on memory systems to model experiences, emotions, and opinions over time. Existing memory benchmarks primarily rely on user-agent conversational histories, which are temporally fragmented and insufficient for capturing continuous life trajectories. We introduce CloneMem, a benchmark for evaluating longterm memory in AI Clone scenarios grounded in non-conversational digital traces, including diaries, social media posts, and emails, spanning one to three years. CloneMem adopts a hierarchical data construction framework to ensure longitudinal coherence and defines tasks that assess an agent's ability to track evolving personal states. Experiments show that current memory mechanisms struggle in this setting, highlighting open challenges for life-grounded personalized AI. Code and dataset are available at https://github.com/AvatarMemory/CloneMemBench

Figures (6)

• Figure 1: Illustrative application scenarios of an AI Clone grounded in long-term digital traces.
• Figure 2: Overview of the Avatar Memory Benchmark data construction and evaluation pipeline. The pipeline proceeds from persona, macro-level life arcs and major events, through meso-level rolling generation of phases with state snapshots, to micro-level detailed event simulation, evidence extraction and final generation of non-conversational digital traces. Extracted evidence is then organized into arc-specific buckets to support evidence-grounded question--answer generation. A detailed version of the pipeline is provided in Appendix \ref{['sec:appendix_A']}.
• Figure 3: Illustrative and representative examples of CloneMem tasks. The left panel shows non-conversational digital traces and their associated ground-truth evidence generated during data construction; the right panel shows example questions and answers for three task types.
• Figure 4: High-level illustration of the CloneMem data construction pipeline. Starting from persona initialization and macro-level life planning, the pipeline expands predefined and model-augmented event seeds into life arcs, major events, phases, and fine-grained detailed events. Each detailed event is generated together with explicit evidence, which jointly ground the generation of non-conversational digital traces. Finally, evidence is aggregated over sliding time windows and life arcs to construct temporally grounded QA instances.
• Figure 5: Dataset composition statistics for CloneMem. Left: distribution of question dimensions (opinion, experience, emotion). Middle: distribution of question types (reasoning categories). Right: composition of media types in the underlying digital traces.