AI Memory Storage Is Cheap, Governance Is More Complex
An analysis published through a WeChat account and authorized versions from TMTPost and 36Kr reports that Alibaba Cloud's PolarDB Agent Memory standard edition charges RMB 0.015 per hour for computing and storage resources for 10,000 memories in mainland China, or RMB 10.8 over 30 days, with model calls billed separately. It contrasts memory retrieval with full-context approaches: in Mem0's 2025 LoCoMo experiment, the memory approach used fewer tokens and had lower latency, while the full-context approach received a higher model-judge score. In a separate Redis LongMemEval test, combining distilled facts with original excerpts produced higher accuracy than using distilled memories alone; the report cautions that the tests used different setups and do not establish a product ranking. It also reviews cross-conversation personalization in ChatGPT, Claude and Grok, and highlights enterprise requirements involving permissions, versions, provenance, updates, expiration, error tracing and portability.
The reports show that low storage fees alone do not determine an AI memory service's enterprise value. Retrieval accuracy, maintenance, rework, authorization, timely updates, error tracing and reliable migration determine whether stored information can continue to support real work.
(本文作者为 冯伟冯大白,钛媒体经授权发布)
文 | 冯伟冯大白
AI记忆的基础费用,已经低到很难与“天价”联系起来。阿里云PolarDB Agent Memory标准版在中国内地,每万条记忆每小时的计算与存储资源费用为0.015元。数量始终保持一万条,30天对应10.8元,模型调用另外计费。
争夺这门生意的公司却越来越多。OpenAI的ChatGPT、Anthropic的Claude和Grok,都已提供跨对话的个性化能力;数据库厂商和独立服务商,则希望把记忆提供给更多应用。Mem0在2025年10月公布种子轮与A轮合计2400万美元融资,同期披露的API调用量,从当年第一季度3500万次增至第三季度1.86亿次。
各家争取的,是用户和企业已经积累下来的东西:反复交代的要求,处理过的异常,否决过的方案,以及做完一半的工作。保存几句话可以很便宜,让这些积累在下一次任务中继续起作用,却会影响用户愿意把工作交给谁,以及是否愿意换一家服务。
#PolarDB Agent Memory#AgentRun#ChatGPT#Claude#Grok#Mem0#LongMemEval#Oracle Agent Memory#Tencent Cloud Agent Memory#Redis#X