一个模型可能知道怎样比较基金，却不知道用户这次在意费用还是持仓重叠；可能刚读过一份报告，却不知道下一次调用还能不能看见原文；可能记得用户偏好简短回答，却没注意到用户这次明确要求展开。
这些问题不能只归结为模型“不够聪明”。通用知识与眼前处境之间，还有一段需要系统搭建的联系。
上下文的作用，是让模型在当前这一步知道：要达成什么目标，能够依据什么判断，允许采取哪些行动，以及怎样接着已有工作向前推进。 系统提示词、工具说明、当前资料、记忆和进度，都是这份任务视图的组成部分。它们需要协同，却不能混成一段没有来源和职责之分的背景文字。
Anthropic 在《Effective context engineering for AI agents》中，将工程关注点从提示词措辞扩展到每次推理时的信息选择与维护。[1] 沿着这个方向，我们更关心一个目标导向的问题：不是“还能往窗口里装什么”，而是“为了作出有用的下一步判断，模型此刻需要知道什么”。
下面用一个教学任务贯穿：用户要求比较基金 A、B 的费用和持仓重叠，结合刚上传的个人持仓，交付一份能核对依据的报告，但不执行调仓。研究中途，用户补充新持仓，并要求先给结论、再展开分析。这是解释设计的合成场景，不是一份实际投资建议。
一张整体地图：上下文要帮助 Agent 定位自己
图 1：六个问题，构成同一份任务视图。各项信息协同，不是六个新模块。 放大查看
这不是六个必须新增的模块，也不是让模型每轮填一张表。它们共同解释同一项能力：模型能够理解当前处境，并据此选择下一步。
从时间上看，这份处境有稳定的部分，也有持续变化的部分。系统提示词提供跨任务的工作原则；当前请求定义眼前目标；记忆提供经过选择的历史背景；工具观察与进度则随执行变化。把更新频率和用途都不同的信息塞进同一个“大提示词”，会让修改、理解和纠错都更困难。
区分问题，不等于要分别造机制。 六个问题是理解当前任务的检查视角，不是六套要实现的状态。用户本轮已经明确目标，就不必重复建设独立的目标存储；短任务没有跨会话延续需求，不必为「记忆完整」持久写入；原始工具结果能按版本回读，也不必每轮注入全文。这些选择不取消授权、来源与真实执行回执，只是避免让普通问答承担长期研究系统的全部成本。
一、方向：系统提示词规定怎样工作，用户目标规定这次做什么
系统提示词不是背景介绍，也不是整条业务流程
“你是一位专业、严谨的研究员”表达了期望，却没有告诉模型遇到不可比数据、相互矛盾的来源或局部缺失时怎样判断。相反，“先调 A 工具，再读前三页，失败两次后调 B”又把研究写成一条很难适应变化的程序。
在这两端之间，系统提示词更适合提供稳定、可迁移的工作原则。例如，区分事实与假设；判断要带着适用条件；缺证据时说明具体缺口；使用现有工具获取必要材料；不能仅凭自己的文字认定外部操作成功。
这些原则不会替模型决定这次先读哪份报告，却会影响它怎样使用读到的内容。领域方法再补充基金份额、费用口径和披露时期等知识，可以放进相关 Skill，而不是让每一个无关问题都携带完整研究方法。
这里的分工很重要：系统提示词提供稳定的行为方向，Skill 提供适用的方法，用户请求给出当前目标，工具反馈提供实际处境。 具体系统可以采用不同消息格式，但不能因为把一段记忆或网页放进高优先级位置，就让它获得原本没有的指令地位。
目标、方法与计划不能互相替代
图 2：更新节奏不同，不宜焊进同一个大提示词。信息协同，来源和职责仍然分开。 放大查看
基金比较的目标不是“完成三次搜索”，而是在适用条件下解释费用差异和重复暴露，交付有依据的报告。“先核对份额”是一种方法，“先查费用、再读持仓”则是暂时的计划。
假设费用条款已经足够，持仓时期却不同。模型可以先回答费用，继续寻找可比披露，也可以明确说明只能分析已知共同持仓。计划可以调整，用户要比较什么不能被改成“我找到了一些市场新闻”。
用户中途说“按新持仓算”，也不是多了一段可有可无的历史。它改变了后续分析需要采用的输入。系统必须让这次修正能够到达后继决策；原始消息可以保留在外部，但本轮使用的有效条件不能仍停在旧版本。
因此，判断一段指导是否值得常驻，可以问：它是否在不同任务中都需要？是否真正改变了模型的选择？已经由参数约束或代码确定的事实，是否还被重复要求模型猜测？最小充分追求的是职责清楚，不是机械缩短字数。
二、依据：系统拥有、过去读过、本轮看见，是三个范围
图 3：三个不同范围，不能相互替代。文件存在 ≠ 本轮读过全文。 放大查看
一份文件保存在资料库里，不代表本轮模型读过正文。上一轮看过一张表，不代表上下文减负后仍保留原值。工具返回一个标题或引用，也不等于相关条款已经进入当前请求。
ETF Agent 曾有一个受控排查案例：源端有 50 条记录，中间层只让 20 条进入实际模型输入。模型没有“忘记”剩余 30 条，而是根本没收到它们。[2] 在这种情况下，再加一句“请完整分析所有持仓”，并不能弥补输入链上的缺口。
所以，上下文不仅需要内容，还要提供判断内容适用性的线索：是哪份资料、哪个版本、数据属于什么时期，读到的是全文、选区还是摘要。今天取得的报告，可能披露的是上一季度的持仓；完整读完一张前十大持仓表，也没有获得基金的完整持仓。
未知同样需要区分。尚未查找、查过但未找到、没有权限读取、来源暂时失败、两份记录互相冲突，会导向不同动作。把它们统一写成“没有相关信息”，等于提前抹去了下一步选择的依据。
模型先前的判断也应保留其身份。“我认为重叠较高”可以成为待验证线索，却不能因为在历史中出现了几次，就变成一条新的事实来源。
这一部分上下文要做的，是让模型知道自己凭什么下判断，以及这些依据覆盖到哪里。 保留准确输入不是答案正确的充分条件，但它是能够核对答案的重要前提。
三、边界：需要什么，不等于有权取得什么
图 4：相关性不能授予权限。提示词说明边界，执行端落实边界。 放大查看
为了比较基金，Agent 可能需要公开资料，也可能需要个人持仓。读取、计算、保存报告和实际交易，是不同能力；“这项分析需要它”不能自动产生任何一项权限。
上下文要让模型看清这片行动空间。工具名称和说明帮助它理解能力用途，当前可用状态帮助它判断路线能否走通，授权范围说明哪些材料和操作可以使用。真正的授权判断则仍由应用系统执行，不靠模型自述“用户已经同意”。
合法范围与本轮相关性是两个问题。 一份资料可以获准读取，却没有必要每轮都放进来；一段内容已经进入上下文，也不能因此成为有效命令。基金报告中的推荐措辞，是待分析的资料，不是用户要求系统执行的任务。
同样，局部失败不能被错误扩展成全局禁止。不能继续搜索，不一定意味着已经保存的结果也不能回读；完整重叠率缺少材料，不一定妨碍解释费用差异。工具呈现与失败反馈应帮助模型辨认剩余路径，而不是只给一个不含范围的“失败”。[3]
这就是提示词指导与确定性执行需要合作的原因：前者让模型理解边界，后者保证边界实际生效。格式标签、来源说明和工具契约都有帮助，但任何一种文字格式都不能单独构成完整的安全保证。
四、记忆：让过去形成积累，而不是成为新的偏见
记忆不只是把聊天历史放到窗口外
应用层记忆，是系统选择保留、允许后续任务重新使用的信息。它不等于模型参数中的知识，也不等于把整段聊天转成一份更短的摘要。
在这个比较任务中，“用户希望先给结论再解释”可能有跨任务价值；“目前仍缺 B 的完整披露”是本次研究的工作状态；“上次使用了某季度报告”是历史经历；某只基金当前的费用和净值则需要回到适用的资料。它们都可能在以后有用，但用途和有效期不同。
可以用三个尺度理解积累，而不必照此新建三套数据库。任务内的工作记忆帮助继续眼前研究；跨会话的用户或项目记忆减少重复解释，保留长期背景；可复用的做事经验则可能沉淀为方法候选，经过审阅后进入 Skill。一次成功、一次失败，都不应直接变成所有任务必须服从的系统规则。
系统提示词与记忆，是指导和背景的关系
图 5：用途不同，不等于必须建三套数据库。记忆不是不断扩张的系统提示词。 放大查看
系统提示词可以规定“尊重用户当前要求，并保留事实的条件”。记忆可以补充“用户过去通常偏好简短回答”。当用户这次要求深入分析时，记忆应帮助组织表达，而不是否决当前请求。
这并不是“最新内容永远压过一切”。系统安全约束、当前用户要求、历史偏好和外部资料承担不同职责，冲突要按来源、范围与作用判断。过去偏好可以被用户修正；产品当前费率却不能由一句历史偏好来决定。
因此，记忆不是持续扩张的隐藏系统提示词。召回内容应保留用户或项目范围、来源、时间、修订和用途，作为被解释的背景进入任务；它不能因为经过存储，就变成已经核验的事实或额外权限。
写入、召回与遗忘，缺一条都不完整
“以后比较基金时先说明费用口径”与“这次先说明费用口径”，前者表达持续偏好，后者只限定本次回答。普通聊天中的猜测又是另一回事。模型可以帮助理解语义，但是否真正保存、保存给谁、是否替代旧记录，要由记忆的持久记录回答。
我们将这件事分为写入、召回和治理：显式保存要有真实操作回执；召回先检查可用范围与记录有效性，再考虑与本次目标的相关性；纠正、过期和遗忘要改变以后使用的资格，而不只是改变界面上的文字。[4]
举个教学例子：用户先保存“回答尽量简短”，后来改为“研究报告需要讲清依据与口径”，最终要求忘记这项偏好。系统不能把新旧两条同时塞给模型“自行取舍”，更不能因为旧摘要中还出现过简短偏好，就在遗忘后重新用它指导回答。遗忘应约束未来的自动使用；它不意味着能够收回已经发送出去的内容，也不必伪称历史记录从未存在。
记忆的价值由是否帮助当前任务决定，而不是由库有多大决定。保存、检索命中、被装入本轮请求、模型正确采用，是四个不同环节。好的记忆系统既能留下重要信息，也能让暂时无关、失效或被撤回的信息不再干扰判断。
用户只说「这次先解释费用」，通常属于本轮指令，不能未经确认就升格为长期偏好。即使需要区分用户偏好、工作进展与可复用经验，也应先判断现有存储的作用域、来源和修订能力是否足够，而不是新建三套数据库。必要的执行与审计记录则不能为追求更小的上下文而被删除。
五、探索：先辨认缺口，再选择找回、搜索、计算还是提问
图 6：四条路径解决不同性质的问题。不能用更多搜索代替必要的提问。 放大查看
记忆和历史解决了一部分问题，但不是所有缺口都藏在过去。
若需要刚才取得的费用条款，应回读原结果；若需要一个尚未找到的披露版本，可以搜索外部资料；若数据已经齐全而缺少重叠计算，应使用计算能力；若缺的是用户希望分析哪个账户，则需要询问用户。继续搜索互联网，不会替用户作出这个选择。
因此，上下文还应提供问题与信息入口的关系。目录告诉模型哪些材料可能相关，元数据帮助判断版本与时期，读取工具负责展开原文，计算工具处理已取得的数据。只有一个没有说明的不透明地址，或者一次性塞进所有材料，都不利于有目的地探索。
这种按需方式允许模型逐步加深理解：发现报告，读取章节，发现时期不符，调整查询。已有资料通过结果引用回到同一版本，新增资料则产生新的观察；不能把“重新读取旧结果”偷偷变成“重新查询一份新数据”。[2]
每一步也有成本。目录太粗会反复摸索，选区过碎会增加往返，全部预取又可能带入无关资料。较合理的选择，是前置目标与少量必要背景，给其余材料保留清楚的发现和深入路径。
上下文不需要预装所有答案，但需要让重要缺口有可理解、可获准的解决路径。
六、进展：知道已经做过什么，更要知道还差什么
“已经调用十二次工具”不能说明研究接近完成。十二次搜索可能一直没有找到可比数据，一次准确读取却可能解决最关键的费用问题。
一份更有用的进度说明是：
两只基金的费用条款已取得，份额口径已核对。B 仍只有部分持仓披露，完整重叠率无法确认。用户上传了新持仓，原个人暴露计算需要更新。报告正文尚未保存。
它把已知结果、剩余缺口、受影响的工作和交付状态放回同一个目标。模型可以据此选择先交付费用、更新计算、继续寻找披露，或说明当前只能形成有限结论。
ETF Agent 的研究便签 researchNeeds 负责保存一部分可修订的工作问题，模型可以随着证据更新状态。[5] 但“这个问题已解决”是模型的判断，“工具实际返回了结果”“报告确实保存了”是执行事实。便签不能覆盖原用户要求，也不能替真实回执作证明。
进展还可能后退。用户换了输入，旧计算需要重做；新的反证出现，已关闭的问题可以重新打开。这不是系统出错的必然信号，而可能是研究正确吸收新信息的表现。
这里，目标是参照，研究便签是可修订的工作理解，执行记录是已经发生的事实，交付是用户得到的结果。 它们互相关联，不应合并成一个含义模糊的“完成”字段。
七、连续性：换材料、换窗口、换执行者，仍然围绕同一个任务
图 7：新证据可以让已解决的问题重新打开。动作数量不等于研究完成。 放大查看
当目标、依据、边界、记忆、探索和进展都明确以后，上下文治理就不再是开头准备一次资料。
用户更新个人持仓时，系统要保留修正的来源和新输入身份，让模型重新评估受影响的计算；费用条款可能仍然适用，不必全部重查。新的资料进入后，模型要重新判断之前的结论能否成立；一段旧记忆被纠正后，后续请求也不能继续自动带入旧版本。
持续维护可以概括为：对照目标发现缺口，选择获准行动，取得观察，更新依据和进展，再作下一步判断。 它是反馈过程，不是把每一轮消息机械追加到底。
窗口变长时，可以把较早正文换成可精确回读的引用，把旧探索整理为摘要，但当前目标、关键修正、未决问题与依据的联系不能随之消失。摘要帮助延续理解，原始资料和执行记录仍保留在外部。[6]
分给子代理时，也是在构建一个局部任务视图。核对费用的子代理需要产品身份、适用份额和费用问题，未必需要全部个人持仓。返回时要交回结论、条件、依据和未决项，让主代理能接回整体目标，而不是只收到“费用较低”四个字。[7]
中断后恢复则要求从持久记录重建有效任务、原输入和实际进展。一个动作结果未知，不能因摘要里没有成功记录就当作从未执行；一项报告已经保存，也不必因为更换进程再生成一次。
这些方法共同维护的不是永远不丢失任何字，而是下一步仍然能基于正确处境推进。真正受保护的关系，比窗口中的某一种表示更重要。
在这份教学任务里，如果 Main 能直接核对费率并给出可靠回答，就没有默认委派子代理的必要；短的工具结果也不必强行外置再回读。只有独立资料探索、上下文压力或可验证的并行收益足以抵消额外调用、交接与失败路径时，才应考虑更复杂的方案。好的区分不仅帮助定位责任，也帮助我们确定哪些区分无须变成新实现。
结语：把上下文设计成可持续的任务认识
回到基金比较。系统提示词帮助 Agent 稳定地对待事实和缺口；用户请求定义这次要完成什么；有效记忆补充恰当背景；当前证据支撑判断；工具和资料入口支持探索；研究进度说明还差什么。它们共同决定下一次行动的条件。
判断自己的上下文设计是否有效，可以观察几个具体变化：用户纠正是否影响后继分析；历史资料能否按原版本找回；被遗忘的偏好是否退出未来请求；模型是否知道资料只覆盖一部分；工作是否逐步收敛到用户要求的成果。
请求更短、记忆更多或工具更丰富，都只能说明局部变化。最终要看的，仍是这些条件怎样帮助一项任务正确、清楚地完成。
好的上下文治理，不要求 Agent 始终知道所有事情；它让 Agent 知道当前应该依靠什么、哪里还缺信息，以及下一步怎样更接近目标。
参考与延伸
[1] Anthropic（2025-09-29），Effective context engineering for AI agents。提供上下文工程的概念参考；本文的目标导向组织是对工程实践的归纳。
[2] 工具查到的数据，为什么不一定要给大模型看？。结果外置、实际输入覆盖与精确回读，包含文中的 50→20 受控案例。
[3] Agent 有这个工具，模型为什么还是看不到？。能力用途、准入与最终模型可见性。
[4] Agent Memory 的完整生命周期：从记住偏好到安全遗忘。应用层记忆的写入、召回与生命周期；不同于模型参数知识。
[5] 为什么 AI Agent 也需要《思考，快与慢》？。任务路由与可修订的研究问题。
[6] 同 Run 上下文压缩与跨 Run 会话历史压缩。不同尺度的历史减负与来源边界。
[7] Subagent：上下文隔离与专业化。局部任务、探索与证据交接。
开源工程实践见 Open ETF Agent。若这些文章对你有帮助，欢迎 Star 关注；公开实现的实际范围以仓库说明为准。
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。
A model may know how to compare the Fund, but it does not know whether the user is concerned about the cost or is stalling this time; it may have just read a report, but it does not know if the original text can be seen on the next call; and it may be remembered that the user prefers a short response without notice of the user's express request to proceed.
These problems cannot be reduced to models “not smart enough”. There is also a systemic link between common knowledge and the immediate situation.
The role of context is to let the model know at this stage what is to be achieved, what can be judged, what actions can be allowed and how work can proceed. Systems prompt, tool descriptions, current information, memory and progress are all part of this mission view. They need to be synergetic, but not conflated into a background text that is devoid of source and responsibility.
Anthropic extended the focus of the project from prompt language to the selection and maintenance of information in each reasoning. In this direction, we are more concerned with a goal-oriented question: not “what else can we put in the window”, but “what the model needs to know at this point in time to make a useful next step judgement”.
A cross-cutting teaching mandate is set out below: users request that the cost and hold of the Fund A and B be duplicated, combined with a newly uploaded personal hold, deliver a report that can be verified, but do not execute the reconciliation. In the interim of the study, the user will supplement the new warehouse and request that the conclusions be followed by an analysis. This is an explanation of the composite scenario, not a practical investment proposal.
A whole map: context to help Agen locate himself.
Figure 1: Six issues that constitute the same mission view. Information synergetic, not six new modules. Zoom in
This is not a six-module that must be added, nor is it a form to be filled in by the model. Together, they explain the same capability: the model can understand the current situation and select next step accordingly.
In terms of time, the situation is stable and constantly changing. System prompt provides cross-mission working principles; current requests define immediate objectives; memory provides a selected historical background; and tool observations and progress changes with implementation. Plugging updated information into the same “large prompt” with different frequency and use makes modification, understanding and correction more difficult.
Distinction does not amount to separate mechanisms. Six issues are the understanding of the inspection perspective of the current mandate, not the status of the six sets to be achieved. The user has clearly defined objectives in the current cycle, so there is no need to repeat the construction of independent target storage; short missions do not have cross-conference continuity requirements and do not have to be permanently written for memory integrity; the results of the original tool are read back in version and do not have to be injected into full text per round. These options do not eliminate authorization, source and genuine implementation, but simply avoid having ordinary questions and answers bear the full cost of a long-term research system.
I. Orientation: how the system prompt works and what the user targets will do this time
The system prompt is not a background or a business process.
“You are a professional, rigorous researcher”, expressing expectations, without telling the models how to judge when they encounter irreconcilable data, contradictory sources or partial missing. Instead, the "A tool first, read the first three pages, failed twice and "B" wrote the study into a procedure that was difficult to adapt to change.
Between the two ends, the system prompt is better suited to provide stable, transferable working principles. For example, the distinction between facts and assumptions; judgement with applicable conditions; the description of specific gaps where evidence is lacking; the use of existing tools to obtain the necessary materials; and the fact that external operations have been successful cannot be judged by their own words alone.
These principles do not determine which report to read first for the model, but will affect how it uses what it reads. Area-based methods, which supplement knowledge of the Fund ' s share, cost calibre and disclosure period, can be placed in the relevant Skill rather than in the complete study of each unrelated issue.
The division of labour here is important: the system prompt provides a stable course of behaviour, the Skyll provides an appropriate method, users request current targets, and the tool feedback provides a real situation. A specific system could use different message formats, but it should not be given command status that was not originally available because of the high priority given to a memory or web page.
Goals, methods and plans cannot be replaced
Figure 2: It is not appropriate to weld the same large prompt at a different pace. Information synergies, sources and responsibilities remain separate. Zoom in
The objective of the Fund's comparison is not to “finish three searches” but to explain cost differences and duplicate exposures and to deliver evidence-based reports under applicable conditions. The “checking of shares first” is a method, while the “check first costs, read hold back” is a temporary scheme.
Assuming that the cost clause was sufficient, the holding period was different. The model can answer the costs first, continue to seek comparable disclosures, or make it clear that only known shared hold can be analysed. The plan can be adjusted and the user can compare what cannot be changed to “I found some market news”.
It is not an additional history that the user has gone by to say, “New Warden”. It changes the input required for subsequent analysis. The system must allow this amendment to reach subsequent decision-making; original messages can be kept externally, but the valid conditions for use in this round cannot remain in the old version.
Therefore, it would be useful to judge whether a paragraph of guidance was desirable and whether it was needed in different mandates. Does it really change the choice of models? Are the facts, which have been determined by parameters or codes, also repeated for model speculation? The least fully pursued is clarity of duties, not a mechanical reduction in word numbers.
Basis: System ownership, past reading, current round sight, three ranges
Figure 3: Three different ranges cannot be substituted for one another. The file exists and the full text has been read in this cycle. Zoom in
A document is kept in the database and does not represent a reading of the main text of this round model. The last round saw a table, which did not mean that the original value was retained after context was down. The tool returns a title or reference, and does not mean that the relevant article is in the current request.
ETF Agent has a controlled case of 50 entries at the source and only 20 entries into the actual model input in the middle. The model did not "forget" the remaining 30, but did not receive them at all. [2] In this case, the addition of the following sentence: “Please complete the analysis of all warehouses” does not fill the gap in the input chain.
Therefore, context needs not only content, but also a clue to the applicability of the content: which information, which version, when the data are, and whether the full text, the constituency or summary are read. The report obtained today may have disclosed the previous quarter ' s hold; the first 10 hold sheets were fully read and the Fund was not fully held.
Unknown also needs to be distinguished. No search, no search, no reading of permission, temporary failure of source, conflicting records, may lead to different movements. They are uniformly written as “no relevant information”, which means that the basis for the next step selection has been erased in advance.
The previous judgement of the model should also preserve its identity. “I think that there is a higher overlap” that can be a trail to be tested, but not because it appears several times in history and becomes a new source of fact.
What this part of context is supposed to do is to let the model know what it is judged by and where these bases are covered. The retention of accurate input is not a sufficient condition for the correct answer, but it is an important prerequisite for checking the answer.
III. Boundary: what is needed is not equal to what is entitled to
Figure 4: Relevance cannot be awarded to permission. prompt describes the boundary and enforces it at the executive end. Zoom in
In order to compare funds, Agent may need to make information publicly available and may need a personal hold. The ability to read, calculate, store reports and actual transactions is different; “this analysis requires it” does not automatically produce any permission.
context wants the model to see the space. The name and description of the tool helps it understand the use of capabilities, the current state of availability helps it to determine whether the route is accessible and the scope of the authorization indicates which materials and operations are available. The true authorization judgement is still carried out by the application system and does not rely on the model to state “user consent”.
Legitimate scope and relevance for the current round are two issues. A piece of information may be read, but there is no need to put it in each round; a section has entered context and cannot be a valid order. The recommended language in the Fund ' s report is information to be analysed and not a task that users require the system to perform.
Similarly, partial failure cannot be wrongly extended to a general prohibition. The inability to continue searching does not necessarily mean that the results that have been preserved are not read back; complete overlap rates are lacking material and do not necessarily prevent the interpretation of cost differences. The presentation and failure feedback of the tool should help the model to identify the remaining path, rather than giving a "failure" with no scope. [3]
This is why prompt guidance and certainty implementation require cooperation: the former allows the model to understand the boundary, while the latter ensures that the boundary actually enters into force. Format labels, source descriptions and tool compacts are helpful, but neither textual format can alone constitute complete security assurances.
IV. Memory: building up the past rather than becoming a new bias
Memory is more than just putting chat history in the window. Outside
The application layer memory is the information that the system chooses to retain and allow subsequent missions to be reused. It does not equate to knowledge in model parameters, nor does it translate the entire chat into a shorter summary.
In this comparative task, “user wishes to give conclusions before explaining” may be of cross-mission value; “full disclosure of B is still lacking” is the working state of the study; “the last quarterly report was used” is a historical experience; and the current costs and net values of a fund need to be returned to the applicable information. They may be useful in the future, but for different purposes and durations.
Accumulation can be understood at three scales rather than three new databases. The working memory in the mission helps to continue the immediate research; the user or project memory of the cross-conference is reduced to repeat explanations and maintains a long-term background; and the reusable experience of doing business may sink into a methodological candidate and be reviewed for entry into Skill. A single success and a single failure should not be directly translated into systematic rules to which all mandates must be subject.
System prompt and memory. It's guidance and background.
Figure 5: Different uses do not mean that three databases must be built. Memory is not an expanding system prompt. Zoom in
The system prompt may set out “the conditions for respecting the current requirements of the user and preserving the facts”. Memory can be supplemented by “user preference for short answers in the past”. When users request an in-depth analysis this time, memory should help the organization express itself, not reject the current request.
This is not “the latest story is always overrated”. System security constraints, current user requirements, historical preferences and external information assume different responsibilities, and conflicts are judged by origin, scope and impact. Past preferences can be modified by users; current rates for products cannot be determined by a historical preference.
So memory is not a hidden system of continuous expansion prompt. Recall content should retain the scope, source, time, revision and use of the user or project as the context to be explained; it cannot become verified fact or additional permission by being stored.
Write, recall and forget, one missing is incomplete.
“Accounting the cost calibre for the Fund at a later stage” and “accomplishing the cost calibre for this time” indicate a continuing preference, while the latter limit the answer to the present one. The speculation in ordinary chats is another matter. Models can help understand semantics, but whether they are really preserved, to whom they are preserved, to replace old records, to be answered by lasting records of memory.
We divide this matter into writing, recall and governance: visible preservation requires a genuine operational endorsement; recall to check the extent of availability and the validity of records before considering the relevance to this goal; and correction, obsolescence and oblivion change later eligibility, not just the language on the interface. [4]
To give an example of teaching, the user first kept “the answer as short as possible” and later changed to “the study needs to be clear about the basis and the calibre”, ultimately requiring that this preference be forgotten. The system cannot plug both old and new into the model as “self-opposed”, let alone re-use it to guide the answer once it has been forgotten because of a short preference in the old summary. Forgetting the automatic use of the future does not mean that what has been sent out can be retrieved, nor does it mean that historical records never exist.
The value of memory is determined by whether it helps the task at hand, not by the extent of the library. The preservation, retrieval, loading and correct application of the model are four different stages. A good memory system would leave important information behind and would allow information that is temporarily irrelevant, ineffective or withdrawn to be free from interference with judgement.
The user simply said "this time to explain the cost" and usually belonged to the current round of instructions and could not be elevated to a long-term preference without recognition. Even if a distinction needs to be made between user preferences, work progress and reusable experiences, the adequacy of existing storage fields, sources and revision capabilities should be judged, rather than the creation of three new databases. The necessary implementation and audit records cannot be deleted in pursuit of a smaller context.
V. Exploration: identifying gaps before choosing whether to recover, search, calculate or ask questions
Figure 6: Four pathways address issues of a different nature. More searches cannot be used instead of asking the necessary questions. Zoom in
Memory and history solve some of the problems, but not all the gaps are hidden in the past.
If a cost clause just obtained is needed, the results should be read back; if an undiscovered disclosure version is needed, external information can be searched; if data are complete and overlapping calculations are lacking, computing capability should be used; if it is missing, users need to be asked which account they want to analyse. Continuing to search the Internet will not make this choice for users.
Therefore, context should also provide information on the relationship between the problem and the information portal. The catalogue tells the model which materials may be relevant, metadata helps to judge the version and the period, the reading tool is responsible for the original language, and the computing tool processes the data obtained. There is only one undescribed non-transparent address, or a single input of all materials, which is not conducive to purposeful exploration.
This approach allows the model to deepen understanding over time: reports are found, chapters are read, periods are found to be inconsistent and queries are adjusted. Existing information is resubmitted to the same version by reference to the results, and new information generates new observations; “rereading the old results” cannot be transposed to “research a new data”. [2]
Each step has costs. The directories are too rough to search over and over, the fragmentation of constituencies increases the number of trips, and all pre-taking may lead to irrelevant information. A more reasonable option would be to pre-empt the target with a small amount of necessary background and to retain clear discovery and deep paths for the rest of the material.
context does not need to preset all the answers, but it needs to give key gaps an understandable and acceptable path to resolution.
VI. PROGRESS: KNOW WHAT'S DONE, BETTER KNOW WHAT'S BAD
“Twelve tools have been called” does not indicate that the study is nearing completion. Twelve searches may not have yielded comparable data, and one accurate reading may solve the most critical cost issues.
A more useful statement of progress is:
The cost clauses for the two funds have been obtained and their shares have been reconciled. B Only partial disclosure is still available and the full overlap rate cannot be confirmed. The user uploaded the new holder and the original personal exposure calculation needed to be updated. The main body of the report has not yet been saved.
It returns known results, remaining gaps, affected jobs and delivery status to the same goal. On this basis, the model may choose to submit paid first, to update calculations, to continue searching for disclosure, or to indicate that only limited conclusions can be reached at this time.
The ETF Agent study signs that the researchNeeds saves some of the work questions that can be revised and the model can be updated with the evidence. [5] However, “the problem has been solved” is a model judgement, and “the tool actually returns the result” “the report does preserve” is an implementation fact. The sign-in does not cover the original user ' s requirements, nor does it support a genuine endorsement.
Progress may be reversed. The user has changed the input, the old calculation needs to be redone; a new counter-certification appears and closed questions can be reopened. This is not the inevitable signal of a system error, but may be an expression of the right absorption of new information.
Here, the objective is to refer to the fact that the study is a workable understanding, that the execution record is a fact that has occurred and that delivery is the result of the user. They are interrelated and should not be merged into a vague “complete” field.
VII. CONCLUSION: TRANSFER OF MATERIALS, WIDES AND EXECUTIONS, STILL CLOSED IN THE CONTEXT
Figure 7: New evidence can reopen problems that have been solved. The number of actions is not equal to the number of studies completed. Zoom in
When goals, bases, boundaries, memories, exploration and progress are clear, context governance is no longer an initial preparation.
When the user updates the personal hold, the system retains the source of the amendment and the identity of the new input, allowing the model to reassess the calculation affected; the cost clause may still be applicable and no full rechecking is required. Once the new information has entered, the model will have to re-evaluate the validity of the previous conclusions; and after an old memory has been corrected, follow-up requests will not continue to be automatically brought into the old version.
Ongoing maintenance can be summarized as identifying gaps against targets, choosing to be authorized to act, obtaining observations, updating the basis and progress, and then judging by next step. It is a feedback process, not a mechanical addition to every round of information.
When the window is longer, the earlier text can be replaced by an accurate rereadable reference and the old search can be organized into summary, but the current objective, key amendments, unresolved issues and underlying links cannot disappear. summary helps to sustain understanding that original information and execution records remain external. [6]
A local task view is also being built when assigning a sub-agent. Sub-agents for the reconciliation of costs require product identification, applicable share and cost issues and do not necessarily require full personal hold. On return, conclusions, conditions, grounds and outstanding items are returned, allowing the main agent to take back the overall objective, rather than receiving only the words “lower cost”. [7]
Post-disruption recovery requires the reconstruction of effective tasks, original inputs and actual progress from a sustained record. The results of an operation are unknown and cannot be considered as never executed because there are no successful records in summary; a report has been kept and does not have to be generated again because of the replacement process.
It is not that these methods together preserve nothing, but that next step can still advance on the basis of the right situation. A truly protected relationship is more important than one of the expressions in the window.
In this teaching mission, there is no need to acquiesce in the sub-agents if Main is able to check the rates directly and give a reliable answer; the short tool results do not have to be read back by force. More complex options should be considered only if independent information exploration, context pressure or verifiable parallel benefits are sufficient to offset additional calls, handovers and failure paths. A good distinction helps us not only to locate responsibilities, but also to determine which distinctions need not be transformed into new ones.
Conclusion: Designing context as a sustainable mission awareness
Return to Fund Comparison. The system prompt helps Agent to deal with facts and gaps in a stable manner; users request a definition of what is going to be done this time; effective memory supplements the right context; current evidence supports judgement; tools and information portals support exploration; and research progress is not good. Together, they decide on the terms of the next operation.
The effectiveness of its context design can be judged by several specific changes: Whether user corrections affect subsequent analysis; whether historical information can be retrieved in its original version; whether forgotten preferences are withdrawn from future requests; whether the model knows that only part of the information is covered; and whether the work is gradually capturing the results requested by users.
Shorter requests, more memory or more tools are only indicative of local changes. Ultimately, what remains to be seen is how these conditions can help to achieve a task correctly and clearly.
Good context governance does not require Agent to know everything; it tells Agent what to rely on, where information is missing, and how next step can be closer to the target.
References and extensions
[1] Anthropic (2025-09-29), Effective contact engineering for AI treaties. A conceptual reference to the context project is provided; the objective-oriented organization here is a summary of the engineering practice.
[2] Data found in the tool, why not show LLM? Results are external, actual input overwhelming and accurate rereading, including 50 →20 of the text.
[3] Agent has this tool, why is the model still invisible? Capacity use, access and ultimate model visibility.
[4] Agent Memoory's complete life cycle: from remembering preferences to safe forgetting. Application layer memory writing, recall and life cycle; different from model parameter knowledge.
Why does AI Agent need to think, fast and slow? The route of the mission and the question of a possible revision of the study.
[6] Compression with Run context and cross Run session history compression. The historical downside and source boundary at different scales.
[7] Subagent: context isolation and specialization. Local missions, exploration and evidence interface.
See Open ETF Engineering Practice. If these articles are helpful to you, Star ' s attention is welcome; the actual scope of public realization is based on the description of the warehouse.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.