更多内存不等于更好，IBM研究指出AI代理需精准检索历史
More memory doesn’t automatically make an agent better. When irrelevant history enters the context w...
IBM研究说，给AI代理加更多内存不一定更好，反而可能让无关历史干扰当前任务。他们发现不同模型对记忆的需求不同，比如gpt-oss-120b模型用精选检索比用完整指南集效果更好，还少用很多令牌。如果你在开发需要持久记忆的代理，这个研究方法值得参考。
IBM研究显示，当无关历史进入上下文窗口时，过去经验会与当前任务信息竞争。在八种模型中，强模型受益于完整指南集，弱模型在紧凑核心和任务特定检索下表现更好，饱和模型则无增益。对于gpt-oss-120b模型，精选检索使任务完成率提升16.1个百分点，仅用5%更多令牌，而注入完整指南集用了51%更多令牌且增益较小。研究建议实用方法：广泛捕获，选择性检索，根据模型和任务调整上下文量。MemSearch采用此方法，结合密集搜索和BM25稀疏搜索进行检索，代理可逐步通过三个层级：搜索结果→扩展记忆→原始转录。
More memory doesn’t automatically make an agent better. When irrelevant history enters the context w...
More memory doesn’t automatically make an agent better. When irrelevant history enters the context window, past experience can start competing with the information needed for the current task. Across eight models, 𝗜𝗕𝗠 𝗥𝗲𝘀𝗲𝗮𝗿𝗰𝗵 found that the right memory “dose” varied with model capability: • Strong models with more headroom benefited from the full guideline set. • Weaker models performed better with a compact core and task-specific retrieval. • Saturated models showed no measurable gain. For gpt-oss-120b, curated retrieval improved task completion by 16.1 percentage points while using only 5% more tokens than the no-memory baseline. Injecting the full guideline set used 51% more tokens and produced a smaller gain. A separate study observed a similar tradeoff across memory architectures. Broad retrieval helped with long-context factual QA, while excessive retrieval reduced performance on sequential decision-making by diverting attention from action-critical context. Taken together, the findings suggest a practical approach to agent memory: Capture broadly. Retrieve selectively. Adjust the amount of context to the model and task. 𝗠𝗲𝗺𝗦𝗲𝗮𝗿𝗰𝗵 follows this approach. Markdown remains the source of truth, while Milvus provides a rebuildable index. Retrieval combines dense search and BM25 sparse search with RRF reranking, and agents can progressively move through three levels of context: Search results → expanded memory → original transcript A useful memory system should provide enough context for the current task, then make deeper history available when the agent actually needs it. If you’re building agents that need durable memory, take a github.com/zilliztech/mem… ps://t.co/OoqhEFvafa Rese huggingface.co/blog/ibm-resea… ps://t.c arxiv.org/abs/2608.15008 ps://t.co/T4zV2dN5sN 💬 0 🔄 0 ❤️ 0 👀 42 ⚡
