Takeaway: A5 is a proactive personal assistant that understands you better the more you use it. Through ongoing interaction, it builds and updates its understanding of you and your projects, applying past preferences, experience, and decisions to new tasks. It also detects changes in connected sources, adjusts ongoing work, and proactively suggests next steps. Get started.

As agents become capable of retrieving information, writing code, and completing complex tasks, our expectations extend beyond individual responses. We want an agent to remember interests we have shared when planning a new trip, understand conclusions reached in another conversation when continuing research, incorporate our feedback on a previous draft when writing again, and notice changes in context so it can offer help at the right time.

We believe that a personal assistant should gradually learn our preferences, goals, and ways of working through repeated interaction, then use that understanding to make appropriate judgments and take action.

Yet most agents today are organized around individual conversations, leaving users responsible for carrying information between them. When starting a new conversation, users must select relevant details from past interactions, provide the information again, explain the background, and restate their preferences—for example, interests they previously mentioned, reasons behind earlier revisions, or writing habits they asked the model to change. Even when this background has been retained, changes in requirements or delivery schedules often still depend on the user to notice what has changed, determine how it affects the original goal, and ask the agent to revisit a plan or offer suggestions.

With these expectations in mind, we designed A5, a proactive personal assistant that understands you better the more you use it. We want A5 to take on three kinds of work:

• Carry past understanding into new tasks. An assistant that relies only on the current conversation requires users to explain the background and repeat their preferences in each new session. A5 accumulates and refines its understanding of users and projects through ongoing interaction, distinguishes lasting interests from temporary trade-offs, and applies relevant experience to new requests.
• Keep ongoing tasks aligned with changing information. When materials are loaded only at the beginning of a task, later analysis may continue to rely on outdated requirements or background until the user notices and asks for a revision. A5 detects changes in connected context, assesses their relevance to the current task, and supplies relevant updates to the agent as it works, allowing it to revisit plans and adjust subsequent actions.
• Identify worthwhile next steps before the user asks again. An assistant that waits for users to initiate every task still requires them to notice changes, assess their impact, and issue instructions. A5 considers new evidence alongside existing goals to identify potential follow-up work, leaving the user to decide whether to proceed.

These capabilities depend on lifelong context and proactivity: an agent needs to maintain an evolving understanding of the user and their work, then use it to decide when and how to help. The following demos show how A5 supports this kind of collaboration in project launches and travel planning.

0x01. What A5 Can Do

Project Launches: Keeping Up with Requirements and Proposing Adjustments

Consider a project to build a resource library. A5 can learn the requirements through earlier discussions. When the requirements document changes and the library shifts from “personal use” to “shared team use,” A5 can detect the change and adjust the proposed access permissions and maintenance responsibilities accordingly.

In another project, a supplier’s delay notice arrives in the workspace after a task has finished. A5 considers the original testing requirements and proactively analyzes the impact on the launch schedule. Before the user asks, it can suggest a revised schedule and wait for confirmation before carrying it out.

Travel Planning: Bringing Previously Discussed Preferences into the Next Trip

While planning a trip to Hangzhou, a user wants to spend more time at the Southern Song Guan Kiln Museum. Later, in Shanghai, the user wants to compare the forms and glaze colors of ceramics from different kilns. As the conversation develops, A5 gains a more detailed understanding of the user’s interest in ceramics.

With that understanding, when the user plans a trip to Beijing and asks which exhibits to see at the Palace Museum and the National Museum of China, A5 can recommend objects and suggest what to look for based on this interest, without requiring the user to explain their preferences again.

Does a choice made during one trip reflect an interest worth retaining over time? Which parts of a project’s schedule are affected by a delay notice? To answer these questions and act on them, an agent needs to organize scattered materials, discussions, and decisions into a coherent understanding of the user and their work, while remaining aware of changes. This calls for lifelong context and proactivity.

0x02. Beyond the Chat Box: Why Lifelong Context and Proactivity Matter

By lifelong context, we mean both a lasting understanding of the user’s preferences and ways of working, and the goals, judgments, and progress of work that continues across conversations. An agent needs to understand the reasons and conditions behind this knowledge so it can determine what still applies and what needs to change. In a long-term literature review, for example, it needs to understand which judgments are supported by evidence, which conclusions remain disputed, and which directions are still unexplored. Each new source must be considered alongside the existing research to determine what it adds, what it changes, and where attention should turn next. Multiple conversations can then advance the same research effort, with the understanding gained in each discussion serving as the starting point for the next.

Proactivity requires an agent to seek out new developments relevant to the user’s goals and decide when to intervene and what help to offer. In the same literature review, the agent automatically tracks the latest papers on the current research topic, proactively downloads them into the workspace, and assesses their significance in light of the existing research. If a new paper supplies crucial evidence or challenges an existing conclusion, the agent can identify judgments that need to be reexamined and suggest updating the relevant analysis in the review before the user asks again. The user decides whether to proceed. Whether a suggestion is warranted depends on the new information’s actual impact on the work; when there is no new information substantial enough to advance it, the agent can refrain from making a suggestion.

Lifelong context and proactivity reinforce each other through collaboration. Existing understanding informs proactive judgments, while the user’s responses may reveal gaps or mistakes that guide later corrections. A5 organizes this process around persistent Context sources, conversations, and memory, allowing the agent to decide what is worth retaining, revising, and using while preserving the evidence needed to check its judgments.

0x03. Inside A5’s Architecture: Supporting Lifelong Context and Proactivity

A5 Harness is built around a ReAct loop. Through plugins and core services, it adds cross-session memory, awareness of workspace changes, and background knowledge maintenance. Together, these mechanisms maintain materials, conversation history, and knowledge in a persistent workspace, bringing past experience and new developments into the current task.

• At the start of a conversation, Memory brings existing understanding of the user and workspace into context.
• During execution, the agent uses the Context index to locate materials, reads original sources and memories as needed, calls tools, and continues reasoning from their results. Diff Engine can supply relevant contextual changes to help the agent adjust subsequent actions.
• After an interaction completes successfully, a memory extraction task can examine new information from the turn and save valuable content for future conversations.

While the system is running, A5 also monitors workspace changes and maintains knowledge even when the user is not interacting with it. During idle periods, Autodream organizes memory and the semantic structure of Context. Workspace Wiki presents accumulated knowledge as pages with source citations that users can check and correct. Proposal Engine considers existing understanding alongside new evidence to decide whether to suggest follow-up work, leaving users to decide in Inbox whether to proceed.

The following sections explain how A5 builds and maintains lifelong context, then turns that understanding into proactive assistance.

Building and Maintaining Lifelong Context

Materials on the same topic may be scattered across directories and sources within a workspace. A5 organizes them through a Context index: Context nodes correspond to individual files or saved webpages, while semantic nodes organize them by topic into a hierarchy that can be browsed level by level. The agent can explore topics first, locate relevant materials, and return to the original sources to verify details.

Memory retains goals, preferences, decisions, and their background as they emerge through collaboration. It is stored in Markdown files and versioned with Git. MEMORY.md holds a consolidated understanding that is relatively stable and useful across tasks, and enters context at the start of a conversation. Detailed notes preserve specific facts and the conditions under which they apply, so the agent can read them as needed. This layered structure keeps the initial background concise while preserving details for tracing individual judgments later. In terms of scope, Workspace Memory stores knowledge, decisions, and constraints specific to the current workspace, while User Memory stores personal background, preferences, and habits that apply across projects.

When a user message arrives, Memory Prefetch uses BM25 in the background to retrieve candidates from both types of memory. A prefetch agent then selects relevant content in light of the current conversation and supplies it to the foreground main agent. The main agent can actively search and read memory or continue working on the task while Prefetch runs in parallel. The prefetch agent reuses the main conversation’s system prompt and conversation prefix to improve prefix cache utilization.

The source index and memory need to evolve throughout collaboration. The main agent can maintain memory while working, and after an interaction completes successfully, a background extraction agent also checks for new information worth retaining. Both need to distinguish temporary requests from lasting preferences in context. If the main agent has already modified memory during the turn, background extraction is skipped to reduce duplicate processing and concurrent edits.

During idle periods, Autodream further organizes the accumulated source index and memory. It adjusts the topic structure in the Context index as materials and areas of interest change. To avoid disrupting the index in use, changes are made in a separate candidate index and published only after structural and version checks pass.

When organizing memory, Autodream uses the Git revision from its previous maintenance pass as a baseline. It examines changed notes and related records, merges duplicate information, corrects or removes outdated content, and updates the overview. This process must preserve the conditions under which judgments apply: records that hold under different conditions can coexist, while unresolved conflicts retain their background and are marked for confirmation with the user when they become relevant again. Autodream also considers existing User Memory to determine which workspace experiences can be useful in other projects and gradually incorporates them into user-level memory. Project-specific requirements remain within their respective scopes.

As records accumulate, memory also needs selective pruning. When the overview needs to be condensed, Autodream first saves useful details to notes and reads them back for verification before updating the overview. Ordinary memories that have not been used in foreground interactions for a long time are archived, removing them from default search and prefetch results. Users can still find and restore this content, and protect memories that are important despite being used infrequently.

Turning Contextual Understanding into Proactive Assistance

A5 continuously detects changes in context connected to the workspace. For ongoing tasks, Diff Engine assesses whether a change is relevant and consults the original source when necessary. New information and its potential implications enter context before the next model call in the ReAct loop, allowing the agent to revisit its plan and adjust subsequent actions.

For work worth pursuing separately, Proposal Engine gathers evidence from contextual changes, completed interactions, and knowledge updates. A Proposal Agent then consults materials as needed and decides whether to make a suggestion. It checks proposal history to exclude work that is completed, already underway, or duplicated. To revisit a suggestion the user previously rejected, it must explain what has materially changed. Suggestions that pass this screening undergo validation, deduplication, and frequency controls before entering Inbox, where the user decides whether to proceed. When there is no worthwhile new work to propose, A5 can remain quiet.

A5 also proactively organizes knowledge worth bringing to the user’s attention, helping them understand projects and make subsequent decisions. Guided by the user’s goals and recent interests, it selects content from materials, discussions, and existing conclusions and organizes it into linked topic pages, forming a user-facing workspace knowledge base: Workspace Wiki. After an interaction completes successfully or Autodream finishes a maintenance pass, Harness can launch a Wiki Agent as needed to add, revise, or reorganize these pages. Pages retain source citations and version history so users can read, verify, and correct them.

Throughout this collaboration, A5 draws on its understanding of the user and workspace to judge which changes need attention, what is worth doing next, and which knowledge is worth presenting. Harness centrally manages background work and the execution of suggestions, subject to the corresponding tool and permission rules. New information and user feedback from this collaboration continue to update memory and knowledge.

0x04. From A5 Today Toward a Long-Term Personal Assistant

Starting with the Harness: Exploring Continuous Learning and Co-Evolution

A5’s modules still need refinement through real, long-term use. What Memory should retain, when it should recall or forget information, how conflicts between memories should be handled, and how proactive suggestions should fit the user’s working rhythm all need to improve through people’s experience with the system. The Context index also needs to connect to more types of context, gradually covering information and events across different tools, applications, and data formats so the agent can understand the user’s work against a more complete background. We plan to open-source A5 on GitHub soon and hope to explore the next generation of personal agents together with the open-source community.

In this release, we introduce A5 Harness, which provides the system foundation for cross-session collaboration, context management, and background execution. We believe that proactivity and lifelong context also require targeted training at the model level. Models need to learn how to apply past experience to new tasks, judge which understanding is worth retaining or revising, and decide when to offer help proactively. Training and evaluation also need to cover the full collaborative process across conversations, making memory maintenance, proactive intervention, and understanding user corrections part of the model’s capabilities. Co-designing the harness and model is one direction we are pursuing.

A longer-term question is whether agents can learn continuously through sustained collaboration and turn experience into more stable, transferable capabilities. A correction can update memory; a repeatedly effective method can become a skill; and accumulated collaborative experience may eventually contribute to model updates. We want to explore how these levels of learning can connect, incorporating new experience while preserving existing capabilities and distinguishing personal or project-specific requirements from transferable methods. This process can also involve the co-evolution of the user, agent, and workspace: users clarify their goals and explore new ways of working, agents adjust their understanding and approach to assistance, and the workspace’s knowledge and processes evolve with them.

Further ahead, we want to explore RSI (Recursive Self-Improvement), allowing agents to use problems revealed through long-term collaboration to help improve their own tools, workflows, memory strategies, harness, and even the models themselves. Agents could propose candidate changes and validate their effects, with the updated system then participating in the next round of improvement. Key questions include how to determine whether a change actually improves collaboration and how to make the improvement process itself more effective. Long-term experience could thereby shape the agent’s understanding, its capabilities, and the system that supports it.

0x05. Start with Your Own Workspace

We will open the source code of A5 soon. For now, you can install A5 technical preview that matches your operating system and processor architecture:

• macOS (Apple Silicon)
• macOS (Intel)
• Linux (AppImage, x86_64; not yet thoroughly tested)

Before getting started, prepare an LLM API that supports the Anthropic Messages or OpenAI Chat Completions format, along with its endpoint, model name, and API key. The model service must support streaming output and tool calling. Web search and webpage fetching use free services by default, so no additional search API key is required. For better search or webpage parsing, you can optionally configure Jina or another supported search API. A5 Harness also supports MCP and Skills, which you can configure as needed.

You will also need Git and Poppler for PDF text extraction. On macOS, you can first install Apple Command Line Tools, which include Git, then install Poppler through Homebrew.

# Install Apple Command Line Tools, which include Git
xcode-select --install

# Install Poppler for PDF text extraction using Homebrew
brew install poppler

On macOS, open the DMG and drag A5.app into Applications. On Linux, extract the package and follow the included INSTALL.md to launch the application. On first launch, follow the setup guide to connect a model and create or select a workspace.

References

• Voyager: An Open-Ended Embodied Agent with Large Language Models
• MemGPT: Towards LLMs as Operating Systems
• MemoryBank: Enhancing Large Language Models with Long-Term Memory
• Proactive Agent: Shifting LLM Agents from Reactive Responses to Active Assistance
• A-MEM: Agentic Memory for LLM Agents
• Sleep-time Compute: Beyond Inference Scaling at Test-time
• Darwin Godel Machine: Open-Ended Evolution of Self-Improving Agents
• Memory in the Age of AI Agents
• OpenClaw - Memory architecture
• Supermemory - Company Brain
• Engram - Understanding a Law Firm through Study

Acknowledgments

We thank the following contributors for their help: Shihan Dou, Xiaoran Fan, Chenhao Huang, Jixuan Huang, Senjie Jin, Chenyang Liao, Jiahang Lin, Shichun Liu, Yujiong Shen, Zhe Sun, Junzhe Wang, Zhiheng Xi, Yajie Yang, Jiazheng Zhang, Ming Zhang, Zhihao Zhang, Enyu Zhou, Yuhao Zhou, and Fangyi Zhu. We also thank Tao Gui, Qi Zhang, and Xuanjing Huang for their support.

A5：主动的个人助手，懂你所想，记你所需，先你一步。

Takeaway：A5 是一个越用越懂你的主动式个人助手。它在长期交互中积累和更新对用户与项目的理解，将过去的偏好、经验和决定用于新任务；也会感知已接入资料的变化，调整正在推进的工作，主动向用户提出后续建议。快速开始。

当 Agent 已经能够检索资料、编写代码并完成复杂任务时，我们对它的期待也逐渐延伸到单次回答之外。我们希望它规划新行程时记得此前表达过的兴趣，接续研究时理解另一个会话里形成的判断，再次写作时吸收我们对上一版的修改意见，主动感知上下文变化，适时主动提供帮助。

我们认为，一个真正的个人助手，应该在一次次交互中逐渐理解我们的偏好、目标与工作方式，并根据这些认识，在之后恰当地做出判断、采取行动。

然而，目前大多数以单次会话为中心的 Agent，仍需要用户承担跨会话的信息迁移工作。开启新会话时，用户要从过去的交互中挑出相关内容，重新提供信息、补充背景、强调偏好，如先前会话里表达过的兴趣、提出修改的理由、要求模型改掉的表达习惯。即使这些背景已经被保留下来，遇到需求更新或交付延期，也常常仍要由用户先发现变化、判断它与原有目标的关系，再要求 Agent 重新检查方案或提出建议。

围绕这样的期待，我们设计了 A5，一个越用越懂你的主动式个人助手。我们希望 A5 承担三类工作：

• 让过去的理解延续到新的任务。 只依赖当前会话的助手，需要用户在新会话中重新交代背景、强调偏好。A5 会从长期交互中积累并修正对用户与项目的认识，区分持续的兴趣和临时的取舍，再将相关经验用于新的请求。
• 让正在执行的任务跟上资料变化。 如果仅在任务开始时载入材料，后续分析就可能继续沿用已经过时的需求或背景，直到用户发现并要求重做。A5 会感知接入的上下文的变化，判断它们与当前任务的关系，将相关更新补充给正在执行的 Agent，使其能够重新检查方案、调整后续操作。
• 在用户再次提问前发现值得推进的工作。 等待用户发起任务的助手，仍需要用户先发现变化、判断影响，再下达指令。A5 会结合新证据与已有目标，判断是否有值得提出的后续事项，交由用户判断是否执行。

这些能力共同依赖于终身上下文与主动性，Agent 需要持续理解用户和工作，再依据这份理解判断何时、如何提供帮助。下面通过项目上线和旅行规划两个 Demo，介绍 A5 如何参与这样的协作。

0x01. 通过 Demo 了解 A5

项目上线：跟上需求变化，主动提出调整

例如，在一个开发资料库的项目中，A5 可以通过此前的讨论了解项目需求。当项目的需求文档发生变更，资料库从“个人使用”变为“团队共用”时，A5 能主动发现需求变化，并相应调整方案中的访问权限和维护分工。

在另一个项目中，当任务结束后，工作空间里收到了供应商的延期通知。A5 会结合原有测试要求，主动分析上线排期受到的影响。即使用户尚未提问，它也能主动提出调整排期的建议，并等待用户确认后执行。

旅行规划：把聊过的偏好用于下一次行程

在规划杭州行程时，用户希望多花时间参观南宋官窑博物馆。到了上海，用户又想进一步比较不同窑口的器形与釉色。随着交流推进，A5 对用户陶瓷兴趣的理解逐渐细化。

带着这种理解，当用户规划北京旅行，询问故宫和国博有哪些推荐展品时，A5 便结合这份兴趣推荐展品和观看角度，无需用户重新交代偏好。

旅行中的一次选择，是否代表值得长期保留的兴趣？一份延期通知，会影响项目中的哪些安排？要回答、解决这些问题，Agent 需要把分散的资料、讨论和决定组织成对用户与工作的整体认识，随时保持对变化的感知。这就需要 Agent 具有终身上下文与主动性能力。

0x02. 聊天框之外：为什么需要终身上下文与主动性

我们所说的终身上下文，既包含对用户偏好与工作方式的长期认识，也包含一项工作在跨会话推进过程中的目标、判断和进度。Agent 需要理解这些认识的理由和适用条件，判断什么可以沿用、什么需要修正。例如，在做长期文献调研时，Agent 需要理解哪些判断已有证据支持，哪些结论仍有争议，哪些方向尚待探索。每读到一份新材料，都需要把它与已有研究对照，判断它补充了什么、改变了什么，以及接下来应该关注哪里。多次会话由此能够共同推进一项研究，每次讨论留下的认识，也能成为下一次深入的起点。

主动性则要求 Agent 围绕用户的目标持续发现新的进展，判断何时介入、提供什么帮助。同样在长期文献调研时，Agent 会自动追踪当前研究话题下的最新论文，主动下载到工作空间，并结合已有研究判断其中是否有值得关注的进展。如果新论文补充了关键证据或挑战了已有结论，Agent 可以在用户再次提问前，主动指出需要重新核查的判断，并建议更新综述中的相关分析，由用户决定是否推进。是否提出建议，取决于新信息对当前工作的实际影响；缺乏足以推动当前工作的新信息时，Agent 也可以不提出建议。

终身上下文与主动性在协作中相互作用：已有理解为主动判断提供依据，用户的回应又可能揭示理解中的偏差，成为后续修正的线索。A5 围绕持久的 Context 来源、会话和记忆组织这一过程，让 Agent 自主判断什么值得保留、修正和使用，同时保留核查依据。

0x03. 深入 A5 架构设计：面向终身上下文与主动性

A5 Harness 以 ReAct 循环为核心，通过插件与核心服务，加入跨会话记忆、工作空间变化感知和后台知识维护机制。这些机制共同维护持久工作空间中的资料、会话历史和知识，让过去的经验与新的变化参与当前任务。

• 会话开始时，Memory 将对用户和工作空间的已有认识带入上下文。
• 执行过程中，Agent 借助 Context 索引定位资料，按需读取原文和记忆、调用工具，并根据工具结果继续判断。Diff Engine 可以补充相关上下文的变化，帮助 Agent 调整后续操作。
• 交互正常结束后，记忆提取任务可以检查本轮的新信息，保存有价值的内容，供后续会话使用。

在系统运行期间，即使没有用户交互，A5 也会关注工作空间的变化、维护知识。Autodream 在空闲时整理记忆与 Context 的语义结构，Workspace Wiki 则将积累的知识呈现为带来源引用的页面，供用户核查和修正。Proposal Engine 结合已有认识与新证据，判断是否提出后续建议，交由用户在 Inbox 中决定是否执行。

下面分别介绍 A5 如何建立并维护终身上下文，以及如何将这份理解转化为主动帮助。

建立并维护终身上下文

工作空间中同一主题的内容可能分散在不同目录和来源中。A5 用 Context 索引组织资料：Context 节点对应具体文件或保存的网页，语义节点按主题组织为可以逐层浏览的结构。Agent 可以先浏览主题，再定位材料、回到原文核查。

Memory 保留协作中形成的目标、偏好、决定及其背景，以 Markdown 文件保存，并通过 Git 管理版本。MEMORY.md 承载相对稳定、跨任务有用的综合认识，在会话开始时进入上下文；详细笔记保存具体事实与适用条件，供 Agent 按需读取。这样的分层让会话开始时的背景保持精炼，也为后续追查具体判断保留了细节。按适用范围，Workspace Memory 保存当前工作空间的知识、决定与约束，User Memory 保存跨项目适用的个人背景、偏好和习惯。

用户消息到达后，Memory Prefetch 会在后台通过 BM25 从两类记忆中检索候选内容，再由预取 Agent 结合当前对话筛选，补充给前台的主 Agent。主 Agent 可以主动搜索和读取记忆，也可以继续推进任务，Prefetch 与前台执行并行。Prefetch Agent 沿用主会话的系统提示词与对话前缀，以提高前缀缓存的利用率。

资料索引与记忆需要随协作持续更新。主 Agent 可以在执行中维护记忆，一轮交互正常结束后，后台提取 Agent 也会检查值得保留的新信息。两者都需要结合语境区分临时要求与长期偏好；如果主 Agent 已在本轮修改记忆，则跳过后台提取，减少重复处理与并发改写。

在空闲时，Autodream 会对积累的资料索引和记忆做进一步整理。它会随资料与关注点的变化调整 Context 索引中的主题结构。为避免影响正在使用的索引，修改先在独立的候选索引中完成，通过结构和版本校验后再发布。

整理记忆时，Autodream 会以上次整理的 Git 版本为基线，检查变化的笔记及相关记录，合并重复信息、修正或移除过时内容，并更新概览。这个过程需要保留判断的适用条件：不同条件下成立的记录可以并存，无法确认的冲突则保留背景、标为待确认，在再次用到时向用户求证。Autodream 还会结合已有的 User Memory，判断哪些工作空间经验能够用于其他项目，逐步纳入用户级记忆。项目中的具体要求则保留在各自范围内。

随着记录增加，记忆也需要有选择地精简。概览需要压缩时，Autodream 会先将有用细节保存到笔记并回读核查，再更新概览。对于长期未用于前台交互的普通记忆，则通过归档将其移出默认检索和预取的范围。用户仍可查找和恢复这些内容，保护低频但重要的记忆。

让上下文理解转化为主动帮助

A5 持续感知已接入工作空间的上下文的变化。对于正在执行的任务，Diff Engine 会判断变化是否相关，必要时回查原文。新信息及其可能的影响，会在 ReAct 循环的下一次模型调用前进入上下文，供 Agent 重新检查方案、调整后续操作。

对于值得单独推进的工作，Proposal Engine 会汇集上下文变化、已完成交互和知识更新等证据，再由 Proposal Agent 按需查阅材料、判断是否提出建议。Agent 会检查建议历史，排除已完成、执行中或重复的事项；重新提出用户拒绝过的建议时，需要说明出现了哪些实质变化。通过筛选的建议还会经过校验、去重和频率控制，再进入 Inbox，由用户决定是否执行。没有值得推进的新事项时，A5 可以不提出建议。

A5 也会主动整理值得用户了解的知识，帮助用户理解项目和作出后续决定。它会结合用户的目标与近期关注点，从资料、讨论和已有结论中筛选内容，组织成相互链接的主题页面，形成面向用户的工作空间知识库 Workspace Wiki。交互正常结束或 Autodream 完成整理后，Harness 可按需启动 Wiki Agent，新增、修正或重新组织相互链接的主题页面。页面保留来源引用和版本历史，供用户查阅、核查和修正。 在这些协作过程中，A5 会依据对用户和工作空间的理解作出判断：什么变化需要处理，下一步有什么值得做，以及哪些知识值得呈现给用户。后台工作和建议的执行都由 Harness 统一管理，并遵守相应的工具与权限规则。协作产生的新信息和用户反馈，也会继续用于更新记忆与知识。

0x04. 从当前的 A5，走向长期的个人助手

从 Harness 出发，探索持续学习与共同进化

A5 的各个模块还需要在真实、长期的使用中不断打磨。Memory 应该保留什么、何时召回或遗忘，记忆之间的冲突怎样处理，以及主动建议怎样适应用户的工作节奏，都需要结合大家的使用体验继续改进。Context 索引也需要接入更多类型的上下文，逐步覆盖不同工具、应用和数据形态中的信息与事件，让 Agent 能够在更完整的背景下理解用户正在做的事情。我们将会于不久后在 GitHub 上开放源代码，希望借助开源社区的力量共同探索下一代个人智能体的范式。

我们这次发布了 A5 Harness，它提供了跨会话协作、上下文管理和后台执行的系统基础。我们相信，主动性和终身上下文能力也需要模型层面有针对性的训练。模型需要学会在新任务中使用过去的经验，判断哪些认识值得保留或修正，以及何时主动提供帮助。训练与评估也需要覆盖跨会话的完整协作过程，让记忆维护、主动介入和对用户纠正的理解成为模型能力的一部分。Harness 与模型协同设计是我们关注的方向。

更长远的问题是，Agent 能否在长期协作中持续学习，把经验转化为更稳定、可迁移的能力。一次纠正可以更新记忆，一种反复有效的方法可以沉淀为技能，长期积累的协作经验还可能参与模型更新。我们希望探索这些学习层次怎样衔接，在吸收新经验的同时保留已有能力，并区分个人和项目中的特殊要求与可以迁移的方法。这一过程也可以是用户、Agent 和工作空间的共同进化：用户逐渐明确目标、探索新的工作方式，Agent 调整理解和帮助方式，工作空间中的知识与流程随之演化。

更进一步，我们希望探索 RSI（Recursive Self-Improvement，递归自我改进），让 Agent 根据长期协作暴露的问题，参与改进自己的工具、工作流、记忆策略、Harness 甚至模型本身，提出候选修改并验证效果，再由更新后的系统继续参与下一轮改进。如何判断一次修改是否真正改善协作，以及改进过程本身如何变得更有效，将是其中的重要问题。长期经验由此有机会同时作用于 Agent 的认识、能力和支撑它运行的系统。

0x05. 从自己的工作空间开始

A5 将于近日开源。目前，你可以安装与你的操作系统和芯片架构匹配的 A5 技术预览版：

• macOS（Apple Silicon）
• macOS（Intel）
• Linux（AppImage，x86_64，尚未经过充分测试）

开始前，请自行准备支持 Anthropic Messages 或 OpenAI Chat Completions 格式的 LLM API，并备好接口地址、模型名称和 API Key。模型服务需要支持流式输出和工具调用。网页搜索和网页抓取默认使用免费服务，无需额外配置搜索 API Key；若希望获得更好的检索或网页解析体验，可以选配 Jina 或其他受支持的搜索 API。A5 Harness 也提供了接入 MCP 和 Skills 的能力，可以自行配置使用。

运行前还需准备 Git，以及用于 PDF 文本解析的 Poppler。macOS 用户可以先安装包含 Git 的 Apple Command Line Tools，再通过 Homebrew 安装 Poppler。

# 安装包含 Git 的 Apple Command Line Tools
xcode-select --install

# 使用 Homebrew 安装 PDF 文本解析工具
brew install poppler

macOS 用户打开 DMG 后，请将 A5.app 拖入 Applications；Linux 用户请在解压安装包后参考包内 INSTALL.md 启动应用。首次启动时，请跟随应用引导连接模型并创建/选择工作空间。

参考资料

• Voyager: An Open-Ended Embodied Agent with Large Language Models
• MemGPT: Towards LLMs as Operating Systems
• MemoryBank: Enhancing Large Language Models with Long-Term Memory
• Proactive Agent: Shifting LLM Agents from Reactive Responses to Active Assistance
• A-MEM: Agentic Memory for LLM Agents
• Sleep-time Compute: Beyond Inference Scaling at Test-time
• Darwin Godel Machine: Open-Ended Evolution of Self-Improving Agents
• Memory in the Age of AI Agents
• OpenClaw - Memory architecture
• Supermemory - Company Brain
• Engram - Understanding a Law Firm through Study

致谢

A5 感谢以下贡献者的帮助：Shihan Dou、Xiaoran Fan、Chenhao Huang、Jixuan Huang、Senjie Jin、Chenyang Liao、Jiahang Lin、Shichun Liu、Yujiong Shen、Zhe Sun、Junzhe Wang、Zhiheng Xi、Yajie Yang、Jiazheng Zhang、Ming Zhang、Zhihao Zhang、Enyu Zhou、Yuhao Zhou、Fangyi Zhu。也感谢 Tao Gui、Qi Zhang 和 Xuanjing Huang 的支持。