Rohan Paul@rohanpaul_ai · X传统LLM在长项目易因有限记忆空间遗忘细节。Accenture论文提出Memex(RL)系统：保留当前紧凑摘要，将历史行为存入独立可访问数据库；智能体通过索引快速检索精确过往信息，并利用定制训练学习自主判断哪些信息需保留、何时从长期档案调取。该方法避免历史过载，保持智能体对当前目标的专注，解决多步复杂任务中的信息丢失问题。论文链接：arxiv.org/abs/2603.04257。
AI agents often forget past work， but this Accenture paper method keeps everything reachable.
Traditional LLMs often forget important details during long projects because their limited memory space forces them to discard old information.
This introduces a system that keeps a compact summary of recent work while storing all past actions in a separate， accessible database.
The agent uses smart indexing to quickly look up exact details from this database whenever it needs to recall a specific past event.
A custom training method teaches the agent to decide for itself which information is worth keeping and when to pull data from its long-term archives.
By saving only the necessary summaries in the active workspace， the model maintains a sharp focus on its current goal without being overwhelmed by a massive history.