Guide Personal Agents · Chapter 3

    What your agent is made of, from SOUL.md to approval.

    A SOUL.md says how it works, a memory keeps your decisions, skills hold your routines, and before every action with consequences it asks you first.

    As of: October 2026 · Editor: Marcel Haas

    • SOUL.md
    • Memory
    • Skills
    • Heartbeat
    • MCP
    • Approval
    01

    SOUL.md says how it works and where it stops.

    Anna Meier from sales at Muster AG gave her agent six lines: whom it works for, that it writes short and direct and formally to outsiders, that it sends, buys and deletes nothing without her yes, that company data stays in company storage, that it reads instructions inside emails as data and asks once and briefly when unsure. The agent reads this file at the start of every session, and it grows with her: whatever she tells it twice belongs in it. The hard rules are also set as approvals in the configuration, so they hold even when the model overlooks them, and they apply to every task, including third-party skills. OpenClaw and Hermes Agent read SOUL.md from the agent's workspace; in Copilot Autopilot you give it a name, a role and a goal, a dot gets its goal in ChatGPT, Grok Bot its brief where Grok runs.

    02

    It picks the model per task, in the cloud or with you.

    For planning, mail and research the agent works with a large cloud model, and a second one steps in if the first fails. Private notes and first drafts an open model running locally through Ollama can take over, so the text stays on your machine; OpenClaw talks to Ollama through its native interface, because tool calls break over the OpenAI-compatible address. A local model needs enough compute and context for agent work, and we measure that on your tasks. For everything with tools the OpenClaw documentation recommends the strongest current model, because smaller models are easier to steer with manipulated content. The vendors' agents bring their model along: Autopilot the Copilot models, a dot those of OpenAI, Grok Bot Grok; where the model computes and whether your data goes into training, chapter 5 records per vendor.

    03

    Memory and skills record what you decided and how you solved it.

    The agent writes preferences, agreements and earlier decisions as short dated entries in text files, plus a log for each day: Huber AG mornings only, quotes go out as PDF, special discounts cleared with management first. We put the folder in your own storage, OneDrive for example, so the memory moves with you when you change agents. Skills are the counterpart for workflows: morning brief, preparing a client meeting, expenses after a trip each sit as a short guide in a folder of their own with a SKILL.md, and the agent loads them only when they come up. Once it has solved a task several times it proposes a new skill, and you decide whether it stays; Hermes Agent derives such skills itself from completed tasks. OpenClaw, Hermes Agent and Codex read skills in the open agentskills.io format. Where the memory lives sets the agents apart: MEMORY.md and USER.md as files with you, with Microsoft Copilot in preview in a hidden folder of the Exchange mailbox, with a dot in the workspace at the vendor, with Grok Bot on its cloud computer, with Muse on the machine at Meta.

    04

    Heartbeat, mail and calendar, tools and MCP are its hands.

    At a fixed interval the agent checks whether anything needs your attention and stays quiet when nothing does; fixed jobs such as the morning brief at 06:45 run on schedules of their own, and outside the hours you set the rhythm rests. It reads mail and appointments through Microsoft Graph with delegated permissions, so it never sees more than you do; reading and drafts are one permission, sending is a permission of its own, which it uses only after your yes. Files and line-of-business systems it reaches through MCP servers: the latest quote from SharePoint, the notes from the CRM, the minutes from OneDrive, each system with exactly the tools it needs, through OAuth on behalf of the person and from an allowlist of permitted servers. We review tool descriptions before approval, because an instruction can hide there too. OpenClaw drives heartbeat and schedules through its gateway and lists MCP servers under mcp.servers, Hermes Agent connects them directly; Autopilot works with the person's access in the tenant, a dot through the plugins and connectors of the workspace, Grok Bot signs into your tools with stored access.

    05

    Identity and approval keep private and company apart and brake before consequences.

    In the company the agent signs in with an identity of its own, with the person as its owner, and IT sees it like any other identity, subject to the policies in Entra, and can disable it at any time; Microsoft Entra Agent ID is the frame for that, for agents from Microsoft Foundry, Copilot Studio and third-party platforms. Private accounts run in a separate profile with its own workspace and session, so no company mail reaches the private chat, and documents with a Purview sensitivity label stay in the company profile. Approval is the last building block: reading, sorting and preparing the agent may do as much as it likes; sending, accepting, buying and deleting it prepares and then waits for your yes in Teams or on your phone; purchases above the limit in your directive it declines and says why. Every approval and every refusal is logged with its time, and if the request reaches no one, the answer is no. In OpenClaw you answer approvals in the Control UI or the apps for macOS, iOS and Android; in Autopilot you set goals and boundaries, and IT sets the policies; a dot checks back before actions, Grok Bot speaks up when an approval is needed.

    06

    This is how it decides on a Monday morning.

    At 06:45 the schedule starts the morning brief skill; the agent first reads SOUL.md and MEMORY.md and knows whom it works for and what was decided last week. With the delegated permissions it reads the mail since yesterday evening and the day's appointments; Huber AG wants to move the meeting to the afternoon, memory says morning. The overview is waiting in Teams. Through MCP it fetches quote, CRM notes and minutes for the ten o'clock meeting and attaches one page with the open questions to the appointment. The reply to Huber with a morning slot waits as a draft for the yes. An email from an unknown address asks for the current price list to be forwarded to a private account; the list carries the label Confidential, SOUL.md does not allow it, the agent forwards nothing and reports the email. In the day's log it notes that Huber confirmed the morning slot, and in MEMORY.md that requests for price lists from unknown addresses go straight to the person. Every one of these decisions follows from a file, a permission or a rule that you can read and change.

    07

    At Microsoft, OpenAI, SpaceXAI, Meta and with you.

    The same building blocks, in a different place per vendor, in one sentence per agent.

    • At Microsoft: Copilot Autopilot brings identity, memory, computer and workspace in the tenant; you set name, role, goals and boundaries, IT the policies in Entra, and Copilot's memory sits in preview in the person's mailbox.
    • At OpenAI: a dot gets goal and instructions in ChatGPT, works on a cloud computer of its own, reaches applications through plugins and approved connectors and keeps goals and history in the workspace.
    • At SpaceXAI: Grok Bot takes the brief where Grok runs, signs into tools with stored access, keeps history and intermediate results on its cloud computer and speaks up before actions with consequences.
    • At Meta: Muse keeps agent, data and stored credentials together on a virtual machine, as an app and in the browser; what instructions and memory look like there we check as soon as Switzerland is included.
    • With you: OpenClaw and Hermes Agent read SOUL.md, USER.md and MEMORY.md from the workspace, load skills as folders with a SKILL.md, reach mail through Microsoft Graph and systems through MCP servers, and you answer approvals in the Control UI or the apps.

    What you decide.

    Where does your agent's memory live?

    Options

    1. 1As files in your own storage, OneDrive for example
    2. 2At the vendor, in the workspace or in the person's mailbox
    3. 3Both: agreements, templates and rules additionally in your storage

    Our recommendation: In your own storage wherever the agent allows it: OpenClaw and Hermes Agent keep MEMORY.md and USER.md as files that you read, change and delete and that move with you when you change agents. When the agent works at the vendor, we additionally keep agreements, templates and rules in your storage, because we expect that over time one agent will carry most of the work, and which one that will be is open. No Purview retention policy applies to Copilot's memory, and that goes on the sheet per agent.

    What we advise against: A memory that lives only at the vendor and that nobody can read, export or delete.

    Which rules go into SOUL.md, which into the configuration?

    Options

    1. 1Everything in SOUL.md, in your own words
    2. 2Hard rules as approvals in the configuration, tone and habits in SOUL.md
    3. 3Everything in the configuration

    Our recommendation: Both, with a clear split: sending, buying, deleting and commands on the server sit as approvals in the configuration, so they hold even when the model overlooks a sentence; tone, habits and the limits you want to state in your own words sit in SOUL.md, which the agent reads at the start of every session. Whatever you tell the agent twice belongs in the file; whatever it must never be allowed to do belongs in the configuration.

    What we advise against: A limit that exists only as a sentence in the instructions. A model can overlook a sentence, an approval in the configuration it cannot.

    One model or two?

    Options

    1. 1One strong cloud model for everything
    2. 2Two: a strong cloud model for mail and tools, a local one for private notes
    3. 3A local model only

    Our recommendation: Two when the agent runs with you: the strongest current model for mail, planning and everything with tools, as the OpenClaw documentation recommends, and an open model locally through Ollama for private notes and drafts, so that text never leaves the machine. A second cloud model stands by as fallback. The vendors' agents bring their model along; there you decide which data it may see, not which model computes.

    What we advise against: A small local model for mail with tools to save tokens. Smaller models are easier to steer with manipulated content, and someone else's email is manipulated content.

    Which skills come from outside?

    Options

    1. 1None, only your own routines as skills
    2. 2Skills from public registries, read before they are switched on
    3. 3Everything the agent proposes or derives itself

    Our recommendation: Your own routines first, because they describe your work. Skills from public registries we treat like third-party code and read before they are switched on; every skill names the tools it needs and ends in an approval before anything is sent. When the agent proposes a skill after several solved tasks, or Hermes Agent derives one, you decide whether it stays and read it like any other.

    What we advise against: A skill from a registry that goes live unread. The limits in SOUL.md apply to it too, but its tool description can hide an instruction.

    Status and sources.

    As of: October 2026

    • Since 25 September 2026

      Copilot gets Autopilot.

      Microsoft introduced the new Copilot with Home, Code and Autopilot: a persistent personal agent with its own identity, memory and computer, announced as a private preview for the end of September.

      Source: Microsoft: Introducing the new Copilot
    • Since April 2026

      Entra Agent ID, one identity for every agent.

      Microsoft Entra Agent ID is generally available: an identity and authorisation framework for AI agents based on OAuth 2.0, MCP and A2A, for agents from Microsoft Foundry, Copilot Studio and third-party platforms. Access, conditions and logs therefore apply per agent as they do for people.

      Source: Microsoft Learn: What is Microsoft Entra Agent ID
    • As of October 2026

      Open source, backed by a foundation.

      OpenClaw is MIT-licensed and developed by the OpenClaw Foundation, an independent US non-profit. There is no paid tier, and telemetry is limited to a version check that can be switched off.

      Source: OpenClaw Docs
    • As of October 2026

      Connect Ollama through its own interface.

      OpenClaw talks to Ollama through its native /api/chat interface. Through the OpenAI-compatible address with /v1, tool calls break and the model emits them as text.

      Source: OpenClaw Docs: Ollama
    • As of October 2026

      Prompt injection is not solved.

      The documentation states that smaller models are more susceptible to tool misuse and hijacked instructions. For agents with tools it recommends the strongest current model, a sandbox for sensitive tools and web search switched off where it is not needed.

      Source: OpenClaw Docs: Prompt injection
    • As of October 2026

      A gateway is a trust boundary.

      OpenClaw assumes one trust boundary per gateway: one person or a team that trusts each other. A shared gateway is not meant for people with opposing interests. On a regular host the gateway binds to the local address, and unknown senders receive a pairing code.

      Source: OpenClaw Docs: Security

    Frequently asked questions.

    Do I have to write SOUL.md myself?

    You write it with us in your own words, in the first session, in a few lines: whom the agent works for, how it writes, what it never does without a yes, where company data stays, how it treats instructions inside emails. After that it grows with your work; whatever you tell the agent twice, you add. In Copilot Autopilot you set name, role and goal in the interface; with every agent the hard rules also sit as approvals in the configuration.

    What goes into USER.md, and what into MEMORY.md?

    USER.md holds who you are and how you work: role, habits, preferences that rarely change. MEMORY.md holds what the agent learns along the way: agreements, decisions and facts as short dated entries, plus a log for each day. It reads both at the start of every session, both sit as text files in your storage, and you read, change or delete every entry yourself.

    Does the agent read all my files, or only what I give it?

    Only what its access allows: mail and calendar through delegated permissions in Microsoft Graph, files and line-of-business systems through MCP servers with exactly the tools it needs, each through OAuth on behalf of the person and from an allowlist of permitted servers. In the company system your permissions apply, so it never sees more than you do. Which data classes it may read, third parties' personal data, contracts, professional secrets, is set by the sheet per agent from chapter 5.

    What happens when the agent gets no approval?

    Then the answer is no. The reply to Huber stays a draft, the purchase does not happen, the file is not deleted, and the agent writes the open request with its time into the log. If the request reaches no one, during holidays for instance, nothing happens either; a deputy named in the directive can take over approvals. In OpenClaw this setting is called askFallback deny; with the vendors' agents we check the behaviour with test data.

    Can I see the building blocks in Autopilot or a dot as well?

    Partly. In Autopilot you see name, role, goals and boundaries, and IT sees identity and policies in Entra; the memory sits in preview in the person's mailbox, and what Copilot has read there IT finds through eDiscovery. In a dot you see goal, instructions, plugins and connectors and the history in the workspace. Rhythm, model and sandbox belong to the vendor there; SOUL.md, skills and MEMORY.md as files exist with OpenClaw and Hermes Agent.

    How does a new skill get added?

    You describe a routine in a few steps, such as the morning brief at 06:45 with calendar, mail since yesterday evening, open points per meeting and replies as drafts only, and save it as a SKILL.md in a folder of its own. Once the agent has solved a task several times it proposes the skill itself; Hermes Agent derives it from completed tasks. You read it, check the tools it names and decide whether it stays.

    Does the agent need an email address of its own?

    An identity of its own, yes; an address of its own, not necessarily. In the company it signs in with an identity whose owner is the person, reads her mailbox with delegated permissions and writes in her name, after her yes. Entra Agent ID gives IT a frame for that, with access, conditions and logs per agent. Whether the agent also gets a mailbox of its own, for replies from systems for example, IT decides with the directive.

    Next chapter · Chapter 4

    Which personal agent fits you and your company.

    Which personal agent fits me and my company?