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。
记忆的价值，最终要在连续工作中兑现。
聊天记录还不是工作经验
AI应用可以保存很多对话，长上下文模型也能一次读入大量文字。但完整历史里既有最终决定，也有被放弃的方案；既有长期偏好，也有一次性要求。下一次任务开始时，模型仍要判断哪些内容有关、哪个版本有效。把材料保存下来，只完成了其中一步。
阿里云AgentRun把长期记忆、会话历史与会话状态分开管理。长期记忆保存跨会话的偏好和信息，会话历史保留完整消息，会话状态保存当前任务的变量和进度。用户说“请继续”时，系统既要知道他一贯的要求，也要知道这次做到哪里，不能把上一回的完成状态套给这一次。
这种差别在项目变更时尤其明显。某个方案经过讨论被否决，历史里仍然会有大量支持它的理由。只按文字相似程度检索，旧方案可能比一句简短的取消决定更容易被找到。系统即使准确找回了原话，也可能接着做一件已经不该做的事。记忆必须保留决定之间的先后和适用范围，才有条件成为经验。
LongMemEval基准设计了500个问题，考察信息提取、跨会话推理、时间关系、知识更新和无法回答时的克制。它的小规模版本，单条测试历史约有40次会话、11.5万Token。找出一个名字，与辨认这个人后来换了职责，是不同的要求。持续使用的AI，需要面对后一种变化。
少读材料，也要少丢条件
把历史反复交给模型，最先遇到的是输入与等待成本。Mem0团队2025年的LoCoMo实验中，完整上下文方案每次查询使用约26031 Token，Mem0记忆方案为1764 Token；总响应时间的第95百分位，分别为17.117秒与1.440秒，即约95%的请求在对应时间内完成。
但完整上下文的模型裁判评分为72.90%，Mem0为66.88%。这组结果呈现的是明确取舍：检索记忆大幅减少了输入和等待，完整历史在该次实验中仍有更高评分。少读材料有收益，也有遗漏条件的风险。
如何兼顾效率和细节，才是更值得投入的部分。Redis研究团队2026年发布的LongMemEval测试中，仅使用提炼记忆，六类任务平均准确率为71.2%；同时检索原始片段与提炼事实，达到86.1%。该次测试包含500个问题，采用模型裁判评分，两种配置的差距为14.9个百分点。
提炼事实适合合并跨会话信息、整理变化，原始片段保留容易被摘要省掉的名字、日期和具体措辞。一次例外当时可能不起眼，下一次却恰好决定任务能否执行。保留原始内容，相当于给后续判断留下回查材料的机会。两项研究使用不同测试与配置，分别回答了少读材料的代价、怎样减少信息损失，不能拼成不同产品的排名。
企业最终要算的，还包括维护与返工。相同要求被后续任务多次调用，前期整理才有机会反复使用；文档改得越频繁，抽取、索引和更新的投入也会增加。缓存能够降低重复输入价格，检索本身也要计算，账单不会只剩下一项存储费。
更大的损失可能发生在答案生成以后。漏掉一个交付条件，接下来需要补问、重做和人工修正，省下的Token会被返工吃掉。采购者真正关心的，是一项工作从交给AI到完成，共花了多少时间与费用。记忆服务只有把这笔总账做得更好，低存储单价才有商业意义。
三家AI都在积累对用户的了解
这些技术选择已经进入日常产品。OpenAI的ChatGPT可以使用历史对话、保存的记忆和自定义指令；在支持的套餐、地区与设置下，还可使用已保存文件及连接应用的内容。它提供记忆摘要，并可在回答下方展示参与个性化的来源，用户能够回看或纠正，但摘要与来源都不是完整清单。
于是，了解用户的材料不再只来自聊天窗口。邮件中的项目变更、保存的工作文件，都可能帮助AI接续任务。来源变多，纠错也更复杂：修改一句总结以后，旧信息仍可能留在某封邮件或某份文档里。用户需要找得到错误来自哪里，服务商则需要处理不同来源之间的关系。
Anthropic的Claude把按主题保存的记忆与历史对话检索分别提供。当前版本可在聊天过程中保存记忆，也可搜索旧对话，检索会显示为工具调用。Chat与云端Cowork可以共用记忆，本地运行的Cowork暂不适用；项目与账号记忆能否互通，取决于相应设置。
这种安排把“记住要求”与“回查原话”连接起来，也让积累有机会随工作流转。用户从讨论转到执行任务，先前交代的文风和项目背景可以继续使用。它与前面的研究方向相通，但产品具备这些入口，不等于每次都能正确提取和调用信息，实际效果仍要在连续任务里检验。
Grok官网也列出了跨聊天记忆，可以记住偏好和过去的对话。X端还有一套个性化机制：在相应授权下，公开资料、帖子、互动和兴趣等信息可以用于个性化，并提供独立控制项。这与实时搜索网页、公开帖子的能力有不同用途。搜索补充外部正在发生的事情，个人记忆帮助系统理解用户此前的要求和决定。
三家的信息来源与控制方式不同，共同目标都是减少用户重复交代的劳动。这里讨论的是产品怎样保存和使用上下文，并非模型把每条个人信息永久写入权重。用户体会到的连续性，来自模型与外围系统共同完成的工作。
这也给独立记忆服务增加了竞争压力。如果需求只是记住职业、文风和常用工具，平台内置功能可能已经够用。第三方继续收费的理由，需要进入更复杂的工作：多个模型能否使用同一份经验，不同成员能否按权限调用，业务变更以后能否及时纠正。偏好保存成为基础能力以后，可靠地复用工作积累才更难替代。
企业需要的是能修正的经验
数据库厂商在这里有自己的位置。企业需要AI记住的信息，往往已经与实际业务相连，更新也发生在业务系统里。
Oracle的Agent Memory产品说明举过一个物流场景：工厂每周需要特定物料，运输人员需要相应资质，交付要在指定时间前完成。这些要求要在多次排程中保留，还要随着业务更新。记住一句话之后，系统必须知道它对应哪家工厂、哪批货物，以及什么时候有效。
如果记忆服务另存了一份要求，业务系统修改后，这份记忆也要跟着变化。否则，AI可能很熟练地安排一条已经过期的流程。Oracle把更新、到期和级联删除列为产品能力；腾讯云Agent Memory则区分个人、智能体与团队记忆，团队共享需要明确授权，默认私有。
一条经验能否复用，因而不能只看内容是否正确。某个客户的特殊要求，如果丢掉归属，就可能被用于其他客户；一次临时例外，如果被保存成团队通用规则，后面的智能体可能一直照做。记忆让正确做法容易延续，也会让错误获得更长的传播时间。
这使权限、版本和出处成为服务的一部分。使用者发现问题时，需要分清原始信息是否错误、提炼是否漏了条件、检索是否选错版本，还是模型理解出了偏差。不同环节需要不同修法。换一个更大的模型，不能保证前面保存与检索的问题一并消失。
能够处理这些关系的记忆服务，才可能成为企业长期使用的基础设施。客户支付的不只是保留多少条信息，还包括更新以后旧规则能及时退出，发生错误可以追查，授权撤回后不再继续传播。工作交得越深，这些管理能力的价值就越难用存储单价衡量。
用户能把记忆带走多少
长期积累会影响平台之间的竞争，但它未必只能留在原地。
Claude已经提供记忆导入、导出。用户可以把其他AI整理出的记忆文本粘贴进去，由Claude提取并保存为条目；自己的Claude记忆也可以导出。导入仍是实验功能，未必能完整吸收全部内容，但一部分偏好与经验已经有了随用户迁移的路径。
这改变了服务商留住用户的方式。一份文风偏好可以复制，几个常用项目的介绍也可以重新输入；单靠保存这些条目，很难建立牢固的长期关系。真正难迁移的，可能是支撑经验生效的整套安排。
文字能够写明某个客户有特殊要求，却不自动附带原始邮件、后来撤回要求的记录、团队授权与正在执行的任务状态。搬到另一套系统后，哪些记录应当被检索，哪个版本优先，谁能看到什么，都需要重新成立。导出完成，不等于同一项工作已经能按原来的方式继续。
这也是独立服务商可以争取的空间。平台希望用连续体验留住用户，跨模型的记忆服务则需要证明：更换模型或应用后，积累仍能可靠使用。企业可以据此把经验保留在自己控制的系统中，再选择适合的模型完成任务。双方竞争的重心，从保存信息转向了如何组织和交付这份积累。
AI记忆的高价值最终落在这些具体结果上：一次纠正能减少后续返工，一次有效处理可以帮助同类任务，一条过期要求能够及时退出决策。保存它们可能只需要很低的基础费用，把它们变成可靠、可更新、可迁移的工作能力，则需要长期服务。用户愿意持续付钱的，是已经被这些积累改变的工作效率。
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
Wen Wei Wei Weng White.
The base cost of AI memory is already so low that it is difficult to relate to the “day price”. The standard version of Alibaba CloudPolarDB Agent Memoory in the Mainland of China cost $0.015 per hour for the calculation and storage of resources per 10,000 memories. The amount remained constant at 10,000, with 30 days corresponding to $10.8, and the model called for additional charges.
More companies are competing for this business. ChatGPT in Open AI, Claude in Anthropic and Grok have provided personalized capacity for cross-dialogue; database manufacturers and independent service providers want to make their memories available for more applications. Mem0 announced a total of $24 million in seed wheel and A in October 2025, while the number of API calls disclosed during the same period increased from 35 million in the first quarter of the year to 186 million in the third quarter.
What they are seeking is what users and businesses have built up: repeated requests, handled anomalies, rejected programmes and done half the work. It would be cheap to keep a few sentences, allowing these accumulations to continue to play a role in the next mission, which would affect the user ' s willingness to give the job to whom and to switch to another service.
The value of memory is ultimately realized through continuous work.
Chat records are not work experience.
AI applications can save many conversations and long context models can read a lot of text at a time. But there are both final decisions and abandoned options throughout history; there are long-term preferences and one-off requirements. At the beginning of the next mission, the model will continue to determine which elements are relevant and which version is valid. The material was saved and only one step was completed.
Alibaba CloudAgentRun manages long-term memory, a session history session, separately. long-term memory saves the preferences and information of the cross-session, saves the full message of the session history, saves the variables and progress of the current task for the session status. When the user says “Please continue”, the system needs to know both what he always requests and where he does it this time, and the last completion status cannot be applied to this time.
This difference is particularly evident when projects change. A programme had been rejected after discussion, and there were still many reasons for supporting it in history. The old scheme could be found more easily than a short decision to cancel, only by reference to text similarities. Even if the system accurately recovers the words, it may go on doing something that is no longer appropriate. The memory must retain the order and scope of application of the decisions before it can become an experience.
The LongMemEval benchmark designed 500 questions to look at the extraction of information, cross-session reasoning, time relations, knowledge updates and restraint when not answered. Its small version, with a single test history of approximately 40 sessions and 115,000 Tokens. Finding a name is a different requirement from identifying the person who later changed duties. Continued use of AI requires facing the latter change.
We don't have to lose the material.
The first thing that happens is input and waiting costs. In the LoCoMo experiment of 2025 for the Mem0 team, approximately 26031 Token was used for each query for the complete context programme, and 1764 Token for the Mem0 memory programme; the 95th percentile of the total response time was 17,117 seconds and 1.440 seconds, i.e. approximately 95 per cent of the requests were completed within the corresponding time.
However, the total context model decision was rated at 72.90 per cent and Mem0 66.88 per cent. This set of results presents a clear trade-off: retrieving memory significantly reduces input and waiting, and complete history still scored higher in the experiment. There is a return to less reading materials and a risk of missing conditions.
How to balance efficiency with detail is a much more valuable part of it. In the LongMemeval test of 2026 by the Redis research team at publish, only memory extraction was used, with an average accuracy of 71.2 per cent for the six categories of tasks, while the original clips and extracting facts were retrieved to 86.1 per cent. The test consisted of 500 questions, which were rated using a model referee, with a difference of 14.9 percentage points between the two configurations.
The extracting fact is appropriate to consolidate the cross-session information and to organize the changes, and the original part is kept with names, dates and specific wording that can easily be saved by summary. An exception might not have been forthcoming at that time, but the next one would have coincided with the implementation of the mandate. The retention of the original content would be tantamount to leaving an opportunity for subsequent judgement to be reviewed. The two studies, using different tests and configurations, answered the cost of less read materials, how to reduce the loss of information and could not be put into different product rankings.
The final calculation for the enterprise includes maintenance and return to work. The same requirement has been repeatedly called upon by the follow-up mission to provide an opportunity for prior processing to be repeated; the more frequently documents are modified, the more input will be drawn, indexed and updated. cache was able to reduce the duplicate input price and the search itself was calculated, and the bill would not have left a single storage fee.
Greater losses may occur after the creation of the answer. A delivery condition was missing, which was followed by additional questions, redactions and manual corrections, and the saved Token was eaten back at work. What is really of interest to the purchaser is the amount of time and cost spent on a job from handing over to AI to completion. The memory service will only be commercially meaningful if the general ledger is better performed and the unit cost of storage is low.
All three AIs are building up an understanding of users.
These technology options have entered the everyday product. The ChatGPT of Open AI can use historical conversations, saved memories and customised commands, and the contents of saved files and connected applications can be used in supported packages, areas and settings. It provides a memory of summary, which can be followed by a demonstration of the source of participation in individualization, and the user is able to look back or correct, but neither summary nor the source are complete lists.
As a result, information about users no longer comes from chat windows alone. Changes to the items in the mail and working papers kept may help AI to continue its mandate. The source is becoming more numerous and the correction is more complex: after a summary sentence has been modified, the old information may still remain in a mail or document. Users need to find out where the error originated, and service providers need to address the relationships between different sources.
Claude of Anthropic provides access to both thematic memory and historical dialogue. The current version can be used to save memories during chats, or to search for old conversations, which are shown as tool call. Chat and cloud-based Cowork can share their memories, and locally operated Cowork is not available for the time being; the interconnectivity of project and account memory depends on the corresponding settings.
This arrangement links the “remember requirements” to the “back-to-back” and provides an opportunity for accumulation to follow workflow. Users have moved from discussions to missions, and the previously communicated text and project background can continue to be used. It is in line with the direction of the previous studies, but the product has these access points, which do not mean that information can be correctly extracted and called on each time, and the actual results will still have to be tested in successive missions.
The Grok network also lists cross-chamber memories to remember preferences and past conversations. There is also a personalization mechanism at the X end: with the corresponding authorization, information such as information, postings, interactions and interests can be used for personalization and the provision of independent controls. This is different from the ability to search web pages in real time and to publish posts. The search complements what is happening externally, and personal memory helps the system understand the user ' s previous requirements and decisions.
The three sources of information differ from the way they are controlled, with the common goal of reducing the duplication of effort by users. The way in which context is preserved and used in the product is discussed here, not the way in which each individual information is permanently written into the weights. The continuity experienced by users is derived from work done jointly with the model and peripheral systems.
This has also put additional competitive pressure on independent memory services. If the demand is simply to remember the profession, the style and the tools commonly used, the built-in functions of the platform may already be sufficient. The rationale for continued third-party charges requires more complex work: whether multiple models can be used in the same experience, whether different members can be called on permission, and whether business changes can be corrected in a timely manner. A reliable recovery of work is more difficult to replace when conservation becomes the foundation.
What businesses need is experience that can be corrected.
Database manufacturers have their place here. Enterprises need information that AI remembers, which is often linked to actual operations, and updates occur in the business system.
Oracle’s Agent Memoory product description shows a logistics scenario: The plant requires specific materials on a weekly basis and the transporter requires corresponding qualifications and delivery is to be completed by the specified time. These requests are to be kept in multiple schedules and updated with operations. With one sentence in mind, the system must know which plant it should be responsible for, which shipment and when it will be effective.
If the memory service maintains a separate requirement, this memory will change as the business system changes. Otherwise, AI may be very skilled in arranging a process that has expired. Oracle lists updates, expirys and cascades as product capabilities; Tencent CloudAgent Memoory distinguishes between individuals, AI agent and team memory, and team sharing requires clear delegation of authority, implicitly private.
An experience can be repeated and therefore cannot be viewed solely as correct. Special requirements of a client may be used for other customers if they are lost; a temporary exception may be maintained in the case of a generic team rule, followed by AI agent. Memory makes the right practice easy to continue and gives the wrongs longer time to spread.
This makes permission, version and source part of the service. Users find it problematic to determine whether the original information is wrong, whether the extraction is missing, whether the search is wrong, or whether the model understands deviations. Different parts of the chain require different legal amendments. A larger model does not guarantee that the question of preservation and retrieval goes away together.
The memory services that can manage these relationships can become infrastructure for long-term use by enterprises. Clients pay more than only a few messages, including the timely withdrawal of the updated old rules, the tracing of errors, and the withdrawal of authorizations without further dissemination. The deeper the work is, the more difficult it is to measure the value of these managerial capabilities in terms of unit storage prices.
How much do users take with them?
Long-term accumulation may affect competition between platforms, but it may not necessarily remain in place.
Claude has already provided memory import and export. Users can paste other AI compiled memory texts, extracted and preserved as entries by Claude; their own Claude memories can also be exported. Import is still an experimental function that does not necessarily absorb all content in its entirety, but some preferences and experiences already have paths to migrate with users.
This has changed the way service providers retain users. A cultural preference can be replicated and the introduction of several commonly used items can be re-inputed; it is difficult to establish strong long-term relationships by preserving these entries alone. What is really difficult to move is probably the set of arrangements that support the entry into force of the experience.
The text can indicate that a client has special requirements without automatic attachment to the original mail, subsequent withdrawal of the request, team authorization and ongoing mission status. Once moved to another system, which records should be retrieved, which version should prevail and who can see what needs to be re-established. Export does not mean that the same job can continue as it was.
This is also the space that independent service providers can seek. While the Platform wishes to retain users with continuous experience, cross-model memory services need to demonstrate that the accumulation can still be used reliably after replacing models or applications. On this basis, enterprises can retain their experience in systems under their control and select suitable models to perform their tasks. The competitive focus has shifted from the preservation of information to how this accumulation is organized and delivered.
The high value of AI memory ultimately falls on these concrete outcomes: one correction reduces subsequent return work, one effective treatment can help the same kind of task, and an outdated requirement can be withdrawn from decision-making in a timely manner. Their preservation may require only low base costs, making them reliable, renewable and transferable, and long-term services. What users are willing to pay on a continuous basis is the work efficiencies that have been cumulatively changed.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.