MemoryForge synthesizes lifelong memories to make LLM agents more human-like

Traditional methods for giving LLM agents human-like personas rely on prompt-based conditioning, injecting static text profiles that often lead to generic behaviors. MemoryForge, a new framework from researchers Bohan Tang and Yiwen Guo, replaces this with memory-based conditioning: it synthesizes an autobiographical memory base from brief target personas, allowing frozen LLMs to dynamically retrieve situation-relevant memories. The framework is grounded in cognitive psychology and consists of three core components: a context generator for socio-historical grounding, a life organizer for developmental coherence toward the target identity, and a multi-resolution simulator that balances broad temporal summaries with high-fidelity episodic experiences.

Experiments on PersonaGym (role-play) and SimulatorArena (user simulation) show that MemoryForge's synthesized memory base enables frozen LLMs to exhibit more human-like behaviors than strong descriptive conditioning baselines, across multiple metrics and various LLM backbones. The work formalizes a new task called customized lifelong memory synthesis and suggests that memory-based conditioning could become a powerful alternative to prompt engineering for creating believable agents. The paper is available on arXiv (2608.00007), and the approach requires no fine-tuning, making it practical for a wide range of agentic applications.