Skip to content
The AI Agent Factory

Chapter 8 on today's products

Approved by
Panaversity ·
23 min read

This page backs 8.8. The lesson teaches the two settings that change how you set up a worker's project: keeping copies current, and memory scope. This page holds the products' own settings, with their sources and the date each was checked.

Products change every few weeks. Check the date on each box before you rely on it, and open its source if the date is old.

The title reads "One blueprint, two project setups." A dark banner reads "Governed knowledge and project instructions, maintained at the source." A table compares Claude from Anthropic and ChatGPT from OpenAI in seven rows. Context window: on Claude, model-dependent, up to 1M tokens on supported paid models, and long chats can be summarized. On ChatGPT, it varies by plan and model, so check the current published limits. Instructions: on Claude, account, project and organization scopes, with organization rules taking priority on conflict. On ChatGPT, project instructions override global custom instructions. Project knowledge: on Claude, project files available across chats, with retrieval for larger knowledge bases. On ChatGPT, chats, files and instructions, with file limits that depend on the plan. Source syncing, in gold: on Claude, Google Docs can stay synced in private projects. On ChatGPT, Drive in projects retrieves on demand and does not sync in advance. Memory scope, in gold: on Claude, project memory, plus account memory unless you turn off Use account memory. On ChatGPT, default or project-only memory, and shared projects use project-only memory. Connected apps: on Claude, connectors access sources using your permissions. On ChatGPT, apps access connected sources, and company knowledge provides citations. Sharing: on Claude, project roles of Can view or Can edit, on supported plans. On ChatGPT, chat access or edit access. A key reads "Gold = settings to check before setup." The foot reads "Same governance. Different settings. Uploads are snapshots. Verify source freshness, access and memory scope," and "Availability depends on plan and workspace settings."

Figure 8.9. The same setup on both AI vendors. Claude's column was verified 18 November 2026, for the classic kind of project, and ChatGPT's 7 October 2026, checked again on 21 November 2026. The gold rows are the differences that change your setup.

The two AI vendors' boxes below cover the same seven settings, in the same order. The open-source box covers the settings its tools match, and how to run them safely.

Anthropic, as verified 18 November 2026.

  • Context window. The context window is everything the model can use for its next reply, counted in tokens, pieces of words. On paid plans, the newest models have a 1M-token (one million token) window. Near the limit of the window, Claude summarizes earlier messages and keeps the full history to refer back to. This summarizing needs a setting that lets Claude run code, Create and edit files (code execution), turned on.12 Long chats use more of your usage limit, the amount of use your plan allows.3
  • Instructions. Account-wide instructions in Settings apply to every conversation.4 Project instructions apply to one project.5 On Team and Enterprise, organization owners can set organization instructions, which take priority where they conflict with a member's own instructions.6
  • Project knowledge. Project files are available in every chat in the project.5 On paid plans, when a project's files grow near the window's limit, it switches to retrieval automatically. Claude then searches the files and reads only the parts it needs.7
  • Synced sources. Google Docs added from Google Drive to chats and projects stay synced to the latest version. In projects, this works only in private projects, ones not shared with others,8 and the content refreshes from time to time when you open the project.9
  • Memory. Claude saves memory as topics you can read, edit and delete in Settings. Each project has its own memory, and never uses another project's. A project also uses your memory from chats outside projects. With memory on, those chats can use what the project saved too. To keep a project's memory fully apart, turn off Use account memory on the project's page.10 That is the classic kind of project. As verified 19 November 2026, a rebuilt kind is in beta on Pro, Max and Ultra. It is rolling out in stages, and is not yet on Team or Enterprise.11 In it, a project is one conversation, and its memory is a set of files that Claude writes as it works. You read, edit and delete them under Project settings, Memory.12 On 18 and 19 November 2026, projects of the rebuilt kind showed no Use account memory setting. When asked, a new project's chat said it remembered nothing from outside the project. A chat outside projects, on the same account, said it did. By default, memory does not keep personal or sensitive topics, unless you turn on Include sensitive topics in memory. Memory can be turned off for one chat, inside or outside a project, before the first message. That chat does not use memory, search your past chats or save anything to memory.10 A second setting, Search and reference chats, lets Claude search past chats on paid plans.10 Incognito chats, outside projects, are not saved to memory.13 On Team and Enterprise, owners decide whether memory is available.10
  • Connectors. Connectors are links from Claude to other apps and services. They reach them with your own permissions. If you cannot open a file, the connector cannot either, unless it was set up with one shared login that many people use.14 As verified on November 20, 2026, you add a connector to your account, in Customize, and choose which connectors each chat may use. On Team and Enterprise, an Owner adds it for the organization first.14
  • Sharing. On Team and Enterprise, a classic project can stay private. It can also be shared with chosen people, with Can view or Can edit access, or be opened to the whole organization.15

OpenAI, as verified 21 November 2026.

  • Context window. OpenAI lists context windows by plan and mode on its pricing page.16 OpenAI's help pages do not say how ChatGPT handles older messages when a chat grows long.
  • Instructions. Custom instructions, ChatGPT's account-wide instructions, apply to every chat in your account.17 Project instructions apply only within the project and override, or take priority over, your custom instructions.18
  • Project knowledge. A project holds chats, files and instructions. The file limit is 5 per project on the Free plan, 25 on the Go and Plus plans, and 40 on the Pro, Business, Enterprise and Edu (education) plans.18
  • Synced sources. In a private project, you can add a Google Drive file or folder as a source by pasting its link. The Google Drive app, a connector, does not sync when added inside a project. It can still search and open files when asked.18 OpenAI says app sync, which indexed content for search, is no longer available. Members connect Google Drive for live access to files instead.19
  • Memory. Saved memories are details you ask ChatGPT to remember. Where that behavior is available, they also include details ChatGPT saves as useful. A second setting, Reference chat history, uses past chats. You can review, correct and delete saved memories.20 A project's own memory shows no list. To stop project memory using a chat, delete the chat or, where settings allow, move the chat to another project.18 Temporary chats do not create or update memories.21 A project can use project-only memory, which ignores saved memories and chats outside the project. Shared projects always use project-only memory.18 A dot, OpenAI's always-on agent, shares memory with ChatGPT both ways. For now, you cannot view, delete or change a dot's individual memories. Deleting the dot starts deleting its context. Files and chats it created are kept.22
  • Connectors. In ChatGPT, apps connect it to services such as Google Drive or Slack. You manage them under Settings, then Plugins.23 On Business, Enterprise and Edu, the Company Knowledge plugin searches connected apps within each user's existing access. When an answer includes citations, you can open them and check.24
  • Sharing. A shared project gives members edit access or chat access.18

Open source, as verified 23 October 2026.

  • Instructions. OpenCode reads standing instructions from files named AGENTS.md. It loads one global file for all your work, then each AGENTS.md in your working folder and the folders above it.25
  • Memory. OpenClaw remembers things by writing plain Markdown files in the agent's workspace, a folder on the machine that runs it.26 Its session transcripts, the records of what was said, are kept in a database file on the same machine.27 Omnara keeps each agent as an event log, a running list of what it did.28 Every step is saved, and that record is the agent's state. So the agent can be picked back up after a crash or restart. If you host Omnara yourself, the record sits in your own database.29 Kept on your own hardware, none of this is the matter's case file.
  • Connectors. OpenCode connects to MCP servers, on your own computer or remote. MCP (Model Context Protocol) is a standard way to connect AI apps to other systems.30 KSoR's served route, which Part V builds, is an MCP server too.
  • Running it safely. Memory kept on your machine still goes to the model's provider when it enters the context, unless the model runs on your own computer.31 So keep anything sensitive out of memory files, as you would with a vendor's memory.

Skills, as verified 18 November 2026 for Anthropic and 21 November 2026 for OpenAI. Skills sit outside any one project, so the boxes above leave them out. 8.5 explains them.

  • Anthropic. Skills work on every Claude plan, with Create and edit files (code execution) turned on.2 You turn them on under Customize > Skills.32 On Team and Enterprise, an owner can give a skill to everyone.33 Partners such as Notion, Figma and Atlassian also publish skills. On the Enterprise plan, an organization can turn on skill scanning, which checks uploaded skills and plugins for malicious content.34
  • OpenAI. Skills are under Plugins, on the Skills tab, on the Business, Enterprise, Healthcare and Edu plans.35 On other plans, keep the procedure in the project instructions.18

Plugins, as verified 18 November 2026 for Anthropic and 21 November 2026 for OpenAI. A plugin is a package of several kinds of setup, such as skills and connectors. Like a skill, it sits outside any one project. Section 8.9 says to read every part inside a plugin before you turn it on. Both AI vendors show you those parts before you add one.

  • Anthropic. A plugin can hold skills, connectors, commands and agents. You add one at Customize > Plugins, on the Discover tab. Select a plugin to read what it contains, then select Add. Anthropic reviews the plugins listed in its directory, but not a plugin you add from another source or upload yourself. Adding a plugin does not connect anything. Its Connectors tab shows each connector it uses, and whether that connector is connected on your account.36 Plugins are for the paid plans: Pro, Max, Ultra, Team and Enterprise.37 Anthropic also lists connectors and plugins on a website, the Claude Marketplace.38 On Team and Enterprise, an owner chooses how members get each plugin the organization offers. It can be available to install, installed by default, or required, so members cannot turn it off. It can also be hidden.39
  • OpenAI. Open Plugins in the ChatGPT sidebar to reach the plugin directory, which every plan has. Whether you can install a given plugin depends on your plan, workspace, role and region. A plugin can hold skills and apps, ChatGPT's name for connectors. Select a plugin and review its description, skills, apps and setup requirements before you select Install plugin. Installing can start each app's sign-in, and it cannot get around the provider's or the workspace's permissions. In a workspace, admins manage plugins in the Admin console. The plugin's settings decide whether members can have it, and each app's settings decide what that app may reach. On Enterprise and Edu, admins can also set plugin access by role.40

Chapter 8 on today's products

Question 1 / 4

0 answered

On ChatGPT, which instructions win inside a project?

  1. How large is the context window on paid Claude plans?, Claude Help Center.↑
  2. Create and edit files with Claude, Claude Help Center.↑↑2
  3. How do usage and length limits work?, Claude Help Center.↑
  1. Understanding Claude's personalization features, Claude Help Center.↑
  2. How can I create and manage projects?, Claude Help Center.↑↑2
  3. Set organization instructions, Claude Help Center.↑
  4. Retrieval augmented generation (RAG) for projects, Claude Help Center.↑
  5. Use Google Workspace connectors, Claude Help Center.↑
  6. Google Drive, Claude Docs.↑
  7. Use Claude's chat search and memory to build on previous context, Claude Help Center.↑↑2↑3↑4
  8. What are projects?, Claude Help Center.↑
  9. Projects, Claude Code Docs.↑
  10. Use incognito chats, Claude Help Center.↑
  11. Use connectors to extend Claude's capabilities, Claude Help Center.↑↑2
  12. Manage project visibility and sharing, Claude Help Center.↑
  13. Pricing, ChatGPT.↑
  14. ChatGPT Custom Instructions, OpenAI Help Center.↑
  15. Projects in ChatGPT, OpenAI Help Center.↑↑2↑3↑4↑5↑6↑7
  16. Admin controls, security, and compliance for plugins and apps, OpenAI Help Center.↑
  17. Memory in ChatGPT, OpenAI Help Center.↑
  18. Temporary chat in ChatGPT, OpenAI Help Center.↑
  19. Dots privacy, security and safety FAQs, OpenAI Help Center.↑
  20. Connected apps in ChatGPT, OpenAI Help Center.↑
  21. Company knowledge in ChatGPT, OpenAI Help Center.↑
  22. Instructions, OpenCode v2 documentation.↑
  23. Memory, OpenClaw documentation.↑
  24. Session, OpenClaw documentation.↑
  25. Machines overview, Omnara documentation.↑
  26. Omnara, README on GitHub.↑
  27. MCP servers, OpenCode v2 documentation.↑
  28. Providers, OpenCode v2 documentation.↑
  29. Use skills in Claude, Claude Help Center.↑
  30. Provision and manage skills for your organization, Claude Help Center.↑
  31. What are skills?, Claude Help Center.↑
  32. Skills in ChatGPT, OpenAI Help Center.↑
  33. Plugins, Claude Docs.↑
  34. Use plugins in Claude, Claude Help Center.↑
  35. Claude Marketplace, Anthropic.↑
  36. Manage plugins for your organization, Claude Help Center.↑
  37. Plugins in ChatGPT, OpenAI Help Center.↑
Owner
Panaversity

On this page