Researchers propose MemoryForge framework to synthesize lifelong memory for human-like LLM agents

Researchers from arXiv propose MemoryForge, a framework designed to synthesize lifelong memory for large language model (LLM) agents to enable more human-like behavior. The work targets a limitation of traditional prompt-based methods, which rely on static textual profiles that often result in generic agent behaviors due to the absence of realistic life memory.

The core innovation is memory-based conditioning, which replaces abstract descriptive profiles with an autobiographical memory base. This allows frozen LLMs to dynamically retrieve situation-relevant memory to guide their actions, drawing inspiration from cognitive psychology.

MemoryForge consists of three components: a context generator for socio-historical grounding, a life organizer to maintain developmental coherence toward the target identity, and a multi-resolution simulator that balances broad temporal summaries with high-fidelity episodic experiences.

The authors formalize the enabling task as customized lifelong memory synthesis and demonstrate the framework’s effectiveness through experiments on PersonaGym for role-play and SimulatorArena for user simulation. The synthesized memory base enabled frozen LLMs to exhibit more human-like behaviors than strong descriptive conditioning baselines across multiple metrics and LLM backbones.

The preprint is authored by Bohan Tang and Yiwen Guo and submitted to arXiv’s Computation and Language (cs.CL) section on June 10, 2026.