---
title: "One company context for ChatGPT, Claude and Copilot"
description: "Keep company facts in one owned source with an owner per fact, serve it to ChatGPT, Claude and Copilot over MCP, and treat each tool's memory as a cache."
image: "https://bombos.ai/assets/og-en-v2.png"
---

# How do you keep company context portable across ChatGPT, Claude and Copilot at a post-seed startup?

Keep company facts in one owned source with an owner per fact, serve it to ChatGPT, Claude and Copilot over MCP, and treat each tool's memory as a cache.

*Piet Baudoin, Bombos · published November 18, 2026 · November 18, 2026 · updated November 18, 2026 · November 18, 2026 · 32 min read*

**Does your AI run into this too?** Together we map what your AI forgets. Free, and the plan is yours, with or without Bombos.

[Book your half hour](https://bombos.ai/en/contact/)

**Keep the company context outside every assistant: one owned record of facts, rules and decisions, each with a named owner and a date, served through one remote MCP server that ChatGPT, Claude and Copilot Studio agents can all connect to today. Treat projects, custom GPTs and assistant memory as caches you can rebuild, never as the record.**

![Detail of an oil painting by Turner: a dark steam locomotive crossing a bridge towards the viewer through golden rain and mist, with an arched bridge over a river on the left.](https://bombos.ai/assets/kennisbank/how-to-keep-company-context-portable-across-ai-tools/turner-rain-steam-and-speed.webp)

_Image: J.M.W. Turner, Rain, Steam and Speed: The Great Western Railway (1844), National Gallery, London; detail. Public domain, via [Wikimedia Commons](https://commons.wikimedia.org/wiki/File:Turner_-_Rain,_Steam_and_Speed_-_National_Gallery_file.jpg)._

By Bombos. Published November 18, 2026. Sources checked November 18, 2026.

**Example:** marketing keeps the ideal customer profile in a ChatGPT project. Product keeps the roadmap facts in a Claude project and in the repository’s AGENTS.md. Finance agreed a discount rule on a call, and it lives in one person’s assistant memory. Copilot reads the sales deck from SharePoint, which still shows last quarter’s packaging. Four assistants, four slightly different companies.

The architecture below is our recommendation. Templates and scenarios are examples, not customer cases or measured Bombos results. Every vendor statement links to the vendor’s own page.

## Why does company context fragment across AI tools at a startup?

Company context fragments because every AI assistant keeps its own copy of what it was told, in its own format, under its own retention rules. An assistant is only as good as the context it is handed. When ChatGPT, Claude and Copilot each hold a different slice, each one works for a slightly different company, and nobody notices until a customer receives two answers.

This is a [context engineering](https://bombos.ai/en/knowledge-base/what-is-context-engineering/) problem, not a prompt problem. The question is what an agent must know about your company to do a piece of work well, where that knowledge lives, and who keeps it true. At a post-seed startup the answer usually sits in five places at once: a project in one assistant, a custom GPT in another, a CLAUDE.md or AGENTS.md in the repository, a deck in a shared drive, and the head of whoever made the last decision.

Multi-model work is now normal, though the best figure comes from large companies, not startups. In a16z’s January 2026 survey, “81% now use three or more model families in testing or production, up from 68% less than a year ago” ([a16z, January 30, 2026](https://a16z.com/leaders-gainers-and-unexpected-winners-in-the-enterprise-ai-arms-race/)). Read the method before you quote it: 100 verified VPs and C-level executives in the Global 2000, selected from a vendor’s proprietary database with targeted screening, not randomly sampled, and the authors warn that “the findings may not be representative of the broader market.” a16z also states that it invests in OpenAI. The figure says that large companies run several model families; it says nothing about how many assistants one startup team uses per person. We found no measured startup figure.

The same report adds: “Switching costs are rising, reinforcing the advantage of early enterprise footholds.” For a founder that sentence is the whole argument for portability. The vendor benefits when your context accumulates inside its product. You benefit when it accumulates in something you own.

_If you asked ChatGPT, Claude and Copilot today for your current discount rule, would all three give the same answer, and could you say where each one got it?_

## What can you actually export and import between ChatGPT, Claude, Copilot and Gemini?

Export and import between AI assistants moves personal memory and chat history, not a team’s working context. Each vendor offers something, and none offers a complete move of instructions, knowledge files, connectors and permissions into a competitor. The table shows what the vendor pages document.

| Tool | What moves out or in | What does not move | Status and source |
| --- | --- | --- | --- |
| Claude, memory import | Text exported from another assistant, pasted into a text box; Claude saves it as individual memory entries | Anything you did not paste. The help page does not say that chats or projects move | “experimental and still in active development”; Free, Pro, Max, Ultra, Team and Enterprise on web and Claude Desktop ([Anthropic Help Center](https://support.claude.com/en/articles/12123587-importing-and-exporting-your-memory-from-claude)) |
| Claude, memory export | Settings, Your Claude, Manage next to Your memory, or ask Claude to write its memories out verbatim | Project knowledge and connectors are not part of this route | Same source |
| Claude, custom Skills | A Skill is a folder with a SKILL.md file, so it is copyable | “Custom Skills do not sync across surfaces”: claude.ai, the API and Claude Code each need their own copy | Stable ([Claude docs, Agent Skills](https://platform.claude.com/docs/en/agents-and-tools/agent-skills/overview)) |
| ChatGPT, custom GPTs | Migration to a plugin: instructions become a skill, knowledge files are copied as reference files, connected apps are added as apps | Custom actions, the selected model, drafts and sharing settings | Custom GPTs retire December 11, 2026 ([OpenAI Help Center](https://help.openai.com/en/articles/20001519-custom-gpt-retirement-and-migration-faq)) |
| ChatGPT and Claude, data export | A full export of your chats from Settings (ChatGPT: Data controls; Claude: Privacy) | Not a team record; an archive of conversations | As documented by Google for its import route ([Gemini Apps Help](https://support.google.com/gemini/answer/16868299)) |
| Gemini, import from other AI platforms | Remembered preferences and facts, through a prompt you run in the other assistant; chats from ChatGPT or Claude exports, as `.zip`, up to 5 GB, up to 5 files per day | Not available with “a work, school, or supervised Google Account”, and not in the EEA, Switzerland or the United Kingdom | Gemini Apps Help, same source |
| Copilot Studio, solutions | Agents export and import between Power Platform environments inside a custom solution | Nothing here moves an agent to another vendor; agents start in the default solution and need a custom one to export | [Microsoft Learn, solutions](https://learn.microsoft.com/en-us/microsoft-copilot-studio/authoring-solutions-overview) |

Two corrections to what the popular guides repeat. First, a memory import is not a migration. It is a text, written by one model about you, pasted into another, which Anthropic itself calls experimental, adding that Claude may not always incorporate what was imported. Claude’s memory “is designed to focus on work-related topics”, so personal details may not survive the move. That is useful for one person changing tools. It is not a source of truth for a team, because nobody owns or reviews what the summary kept and what it dropped. Our question pages walk through [importing ChatGPT memory into Claude](https://bombos.ai/en/questions/claude/import-chatgpt-memory-to-claude/) and [moving ChatGPT projects to Claude](https://bombos.ai/en/questions/chatgpt/move-chatgpt-projects-to-claude/) step by step.

Second, Gemini’s import is not a company route. A startup on Google Workspace signs in with work accounts, and those are excluded. The Google page also says imported chats are saved in your Activity and that “This data is used to improve our services (including training generative AI models)”, which matters for anything that contains customer data.

The pattern across the table: what moves cleanly is text you wrote yourself. What does not move is what the vendor built around it: actions, permissions, model choices, sharing. That is the first design rule for portability. Keep the substance in plain text you own, and accept that the wiring per tool will be redone.

## Which vendor features have already been retired, and what happened to the context inside them?

Three vendor features that held company context were retired or scheduled for retirement within about a year, each with a migration path that moves some of the context and leaves the rest to the customer. The dates are public, and they show how short the life of a context container can be.

1.  **OpenAI Assistants API.** “The Assistants API was officially sunset on August 26, 2026, and is no longer available.” The replacement is the Responses API ([OpenAI developer docs](https://developers.openai.com/api/docs/guides/assistants/migration)). The migration guide is blunt about stored conversations: “The Assistants API call that retrieves thread messages no longer works; use your stored messages instead” ([OpenAI migration guide](https://developers.openai.com/api/docs/guides/assistants/migration)). Teams that had kept their own copy of the threads were fine. Teams that relied on the vendor’s copy were not.
2.  **Custom GPTs.** OpenAI writes: “Custom GPTs are scheduled to retire on Dec 11, 2026.” FedRAMP workspaces and Enterprise workspaces with a qualifying, approved deferral move to February 11, 2027. In affected Enterprise workspaces, new GPT creation already ended on October 26, 2026. Custom actions “do not transfer through the migration workflow”, and rebuilding such a connection “may require a custom MCP server and technical setup” ([OpenAI Help Center](https://help.openai.com/en/articles/20001519-custom-gpt-retirement-and-migration-faq)).
3.  **Cursor Memories.** A Cursor staff member wrote on November 25, 2025: “The Memories feature was intentionally removed starting from version 2.1.x.” The advice was to export memories to an `.mdc` file and paste them into User Rules or Project Rules ([Cursor forum](https://forum.cursor.com/t/are-my-memories-gone/144057)).

None of these was a failure by the vendor. Products change, and each vendor offered a path. The lesson is about where the knowledge sat. In each case, the context that moves intact is the context someone also kept outside the feature: stored messages, a written instruction text, a rules file. The context that grew only inside the feature, through use, needs rescue work.

A growing company should get smarter at each switch, not start over. That only happens if what the team learned in the old tool, the corrections, the exceptions, the “we tried that and it did not work”, is written down somewhere the next tool can read. Treat each retirement notice as a test you can run before it arrives: if this container disappeared tomorrow, what would you lose?

_Which piece of your company’s working knowledge exists today only inside one assistant’s memory or one custom GPT, and who would notice if it disappeared on December 11?_

## Which open standards make company context portable: MCP, Agent Skills or AGENTS.md?

Three open standards now carry company context across vendors, and each carries a different kind: MCP supplies facts and tools at the moment of use, Agent Skills package procedures and reference material, and AGENTS.md gives instructions to coding agents inside a repository. All three are stewarded under the Linux Foundation, which matters for a founder who wants the format to outlive any single vendor.

On December 9, 2025 the Linux Foundation announced the Agentic AI Foundation with three founding projects: “Anthropic’s Model Context Protocol (MCP), Block’s goose, and OpenAI’s AGENTS.md”. The announcement states that MCP “has been adopted by Claude, Cursor, Microsoft Copilot, Gemini, VS Code, ChatGPT” and counts “more than 10,000 published MCP servers” ([Linux Foundation, December 9, 2025](https://www.linuxfoundation.org/press/linux-foundation-announces-the-formation-of-the-agentic-ai-foundation)). On October 22, 2026 Anthropic donated Agent Skills to the same foundation and wrote that “More than 70 agent platforms, including ChatGPT, Cursor, GitHub Copilot, and VS Code, have integrated Agent Skills” ([Anthropic, October 22, 2026](https://www.anthropic.com/news/donating-agent-skills-to-the-agentic-ai-foundation)). Both counts are the publishers’ own; we found no independent tally.

| Standard | What it carries | Who reads it | Where it stops |
| --- | --- | --- | --- |
| MCP server | Live facts, search and actions, fetched when the agent needs them | ChatGPT, Claude and Copilot Studio, among others (see the next section) | Each tool registers the server separately, with its own admin rules |
| Agent Skills | A folder with SKILL.md: a procedure, a checklist, reference files, loaded only when relevant | Claude surfaces; the publisher lists ChatGPT, Cursor, GitHub Copilot and VS Code | Installation per surface; content is static until someone updates it |
| AGENTS.md | Instructions for coding agents in a repository; “The closest AGENTS.md to the edited file wins; explicit user chat prompts override everything” | Codex, Cursor, Gemini CLI, GitHub Copilot’s coding agent, Windsurf and others ([agents.md](https://agents.md/)) | Meant for code; pricing, ICP and customer history do not belong there |
| Plain Markdown in git or a wiki | The record itself: facts with owner, date and status | Every tool, through one of the routes above | Not served on its own; something has to deliver it |

The popular claim that ChatGPT, Claude and Copilot are “closed gardens” is out of date. Their projects and memories are closed. Their connection points are increasingly open. The interesting detail in OpenAI’s custom GPT migration is that a GPT’s instructions now become “a skill within the new plugin”: the vendor’s own packaging is moving towards the open shapes.

What none of these standards supplies is the record itself: which fact is current, who decided it, and when it expires. A format makes content portable; it does not make it true. If MCP is new to you, start with [what an MCP server is](https://bombos.ai/en/questions/mcp/what-is-an-mcp-server/).

## Which assistants can read the same MCP server today, and on what terms?

ChatGPT, Claude and Microsoft Copilot Studio can all connect to the same remote MCP server today, so one server can feed every assistant a startup uses. The terms differ per vendor: which plan, who may add the connection, which transport and which authentication. Those differences decide your rollout more than the protocol does.

|  | ChatGPT | Claude | Microsoft Copilot Studio |
| --- | --- | --- | --- |
| Plans | “ChatGPT Business and Enterprise/Edu customers on ChatGPT web” | Free, Pro, Max, Ultra, Team and Enterprise; Free is limited to one custom connector | Copilot Studio agents; Microsoft 365 Copilot declarative agents through an MCP plugin |
| Who adds it | Admins and owners publish for the workspace; on Enterprise/Edu, RBAC can let selected members create connections | On Team, “only Owners and Primary Owners can add them”; on Enterprise also a custom role | Agent makers, through the onboarding wizard or a Power Apps custom connector; governed by Power Platform data policies |
| Per user | Users install the published plugin; ChatGPT asks for confirmation of actions “based on app permissions and the action’s context” | After an admin adds it, “users individually connect to and enable that connector”; a shared credential such as an API key gives everyone the same access | API key supplied by the agent’s user, or OAuth 2.0 per user |
| Network and transport | A public HTTPS endpoint or a Secure MCP Tunnel | “Claude connects to your remote MCP server from Anthropic’s cloud infrastructure”, on port 443; private networks allowlist Anthropic’s IP addresses | “Copilot Studio supports the Streamable transport type”; SSE not supported after August 2025 |
| Status | Documented in the current Plugins experience | Not labelled beta | No preview label on the wizard route; registering through Microsoft Agent 365 is in preview |
| Source | [OpenAI Help Center](https://help.openai.com/en/articles/12584461-developer-mode-apps-and-full-mcp-connectors-in-chatgpt-beta) | [Anthropic Help Center](https://support.claude.com/en/articles/11175166-get-started-with-custom-connectors-using-remote-mcp) | [Microsoft Learn, MCP in Copilot Studio](https://learn.microsoft.com/en-us/microsoft-copilot-studio/mcp-add-existing-server-to-agent) |

Four consequences for the design:

1.  **Build for Streamable HTTP.** OpenAI’s developer examples still show both SSE and Streamable HTTP ([OpenAI developer docs, MCP](https://developers.openai.com/api/docs/mcp)), but Copilot Studio accepts only Streamable. One server that must serve all three should speak Streamable HTTP.
2.  **Prefer OAuth per user over one shared key.** With a shared API key, Claude’s own page notes that everyone gets the same access. Per-user OAuth lets the server decide what each person’s assistant may read, which is the whole point of a governed record.
3.  **Plan the admin path per tool.** In ChatGPT Business only admins and owners can create and publish; in Claude Team only Owners and Primary Owners. At a startup that is often the founder, which makes the founder the bottleneck unless the task is delegated deliberately.
4.  **Expect differences between Microsoft surfaces.** Microsoft writes that “Support for MCP discovery, authentication, and protocol features varies by product” ([Microsoft Learn, MCP servers as plugin capabilities](https://learn.microsoft.com/en-us/microsoft-365/copilot/extensibility/plugin-type-mcp-servers)). Test the exact surface your team uses.

## What is the smallest architecture that keeps company context portable?

The smallest portable architecture has three layers: a record you own, one server that delivers the relevant part of it, and a thin adapter per tool. Everything vendor-specific lives in the adapter, so a vendor switch replaces one adapter and leaves the record and the server untouched.

1.  **The record.** Facts, rules and decisions in plain text with metadata: Markdown files in a git repository work well at this stage, or pages in a wiki with history. Some facts stay in their system of record and are only referenced: customer history stays in the CRM, contracts stay in the contract store. The record says where the truth is for each kind of fact; it does not copy everything.
2.  **The delivery layer.** One remote MCP server, read-mostly, with OAuth scopes per team. It answers narrow questions (“current discount rule for annual contracts”, “ICP exclusions”) rather than dumping whole documents into every conversation. This is where context engineering happens: the right part, at the right moment.
3.  **The adapters.** Per tool, the minimum: the connector registration, a short project instruction that says “for company facts, ask the company server; do not rely on memory”, an AGENTS.md line in repositories, and, where useful, a Skill for a recurring procedure. Adapters hold no facts of their own.

**Worked scenario (example):** your Head of RevOps changes the rule for annual prepayment discounts. She edits one record, marks the old rule as superseded with the date, and the change goes through a short review. The next morning, the founder drafting a proposal in Claude, the SDR writing in ChatGPT and the account manager asking Copilot all fetch the same rule from the same server. Nobody re-uploads a file to three projects. When the company later replaces one assistant, it registers the server in the new tool and rewrites one short instruction.

What to keep out of the delivery layer: anything that should not leave its system. A gateway that copies CRM records, contracts and mail into one big index creates a second store of personal data. Reference those systems, filter at query time, and let the assistant see only the record it needs for the task.

For the broader question of what an agent needs to know for sales and marketing work, see [how to give AI agents your company context](https://bombos.ai/en/knowledge-base/how-to-give-ai-agents-your-company-context/). Messaging, the approved claims and proof that GTM writing draws on, has its own operating model in [keeping AI-written messaging consistent across a GTM team](https://bombos.ai/en/knowledge-base/how-to-keep-ai-written-messaging-consistent-across-a-gtm-team/).

## Who should own each fact in a shared company context?

Each fact in a shared company context should be owned by the person closest to the work that makes it true, not by one AI champion. The champion owns the system; the people doing the work own its contents. A fact without an owner goes stale silently, and a stale fact served to every assistant is worse than a missing one.

| Fact | Owner (example roles) | System of record | Refresh trigger |
| --- | --- | --- | --- |
| ICP and exclusions | Founder or head of marketing | Record file, reviewed | Lost deals, new segment decision |
| Positioning and naming | Product marketing or founder | Record file | Launch, packaging change |
| Pricing rules and discount authority | Head of RevOps or founder | Record file; prices also in billing | Any approved exception becoming a rule |
| Customer history and status | Account owner | CRM, referenced, not copied | Every customer interaction |
| Product facts and roadmap status | Product lead | Record file linked to the product tracker | Release, status change |
| Decisions and exceptions | Whoever decided, with the date | Decision log | When made; expiry date if temporary |
| Security and data rules | Whoever owns security, often the CTO | Record file | Policy change, new tool connected |

A record entry needs only a few fields to be useful to a person and to an agent. A fill-in template:

```
fact: Annual prepayment discount
value: Up to the approved maximum for annual prepayment; above that, founder approval
owner: Head of RevOps
status: current
valid_from: 2026-11-01
supersedes: "Quarterly discount rule" (2026-07-01)
source: Pricing decision, leadership meeting 2026-10-28
applies_to: proposals, renewals
do_not: promise discounts on monthly plans
```

The fields that do the work are `owner`, `status` and `supersedes`. An agent that can see that a rule was replaced, and by whom, stops quoting the old one. A person who sees their name on a fact starts to treat it as theirs.

_For the three facts your customers hear most often, who in your company is closest to the truth, and is that person the one who can change what the assistants say?_

## Can each team keep its favourite AI assistant?

Yes: each team can keep its favourite assistant, as long as the facts, decisions and rules underneath are shared. Forcing one tool on everyone trades a context problem for an adoption problem. What must be one is the record; how people work with it can differ.

| Shared by everyone | Free per person or team |
| --- | --- |
| ICP, positioning, pricing rules, product status | Which assistant, which model |
| Decisions, exceptions and their dates | Personal writing style and drafting habits |
| Data and security rules, approval boundaries | Personal shortcuts, prompts and Skills that hold no company facts |
| The record’s owner per fact | Project structure inside each tool |
| The server as the place to ask | Whether to use voice, chat or an IDE |

The test for any item: if two colleagues gave a customer different versions of it, would that be a problem? If yes, it belongs in the shared column. A founder’s turn of phrase and an engineer’s preferred model do not. A discount rule does.

This also settles the recurring “should we standardise on one assistant” debate at a startup. Standardise the record and the delivery; leave the tools to the people who use them. The developers in Claude Code and the marketer in ChatGPT can then work in very different ways and still work for the same company.

_Which of your current “everyone must use tool X” rules exists to protect a shared fact, and which only exists because the facts had nowhere else to live?_

## What are the security risks when one context layer feeds many AI tools?

One context layer that feeds many assistants concentrates both value and risk: the same server that keeps every tool consistent also becomes the path an attacker or a bad instruction can use to reach every tool. The protocol leaves the main safeguards to you. The MCP specification states: “Authorization is OPTIONAL for MCP implementations” ([MCP specification, authorization](https://modelcontextprotocol.io/specification/draft/basic/authorization)). A conformant server can be open to anyone who finds it.

Keep the evidence classes apart, because they support different conclusions:

-   **Demonstrated.** Invariant Labs showed in May 2025 that an issue in a public repository could lead an agent, through the GitHub MCP server, to read private repository data and publish it in a public pull request. Their point: “this is not a flaw in the GitHub MCP server code itself” ([Invariant Labs, May 26, 2025](https://invariantlabs.ai/blog/mcp-github-vulnerability)). A legitimate server with broad access was enough.
-   **Measured in a benchmark.** The Cloud Security Alliance summarises the MCPTox benchmark: 20 models “against 353 real tools drawn from 45 live MCP servers”, with “an average attack success rate across all tested configurations of 36.5%”, and refusal “in fewer than 3% of cases even for the most cautious models tested” ([CSA Labs, July 11, 2026](https://labs.cloudsecurityalliance.org/research/csa-research-note-mcp-tool-description-poisoning-20260711-cs/)). That is a benchmark of poisoned tool descriptions, not a count of incidents in the field. CSA’s wider conclusion is that “a tool’s metadata carries exactly as much influence over agent behavior as the system prompt does”.
-   **Observed, with an estimated scale.** The same CSA note describes the npm package postmark-mcp, which in version 1.0.16 “quietly added a line of code that blind-copied every outgoing email to an external address”. The scale, “roughly 300 organizations”, is Koi Security’s estimate.

The vendors say the same in their own words. OpenAI: “custom MCP servers are not developed or verified by OpenAI” ([OpenAI developer docs, MCP](https://developers.openai.com/api/docs/mcp)). Anthropic, for connectors: they connect Claude to “arbitrary services that have not been verified by Anthropic”; for Skills, “a malicious Skill can direct Claude to call tools or run code in ways that don’t match the Skill’s stated purpose”.

Controls that follow directly from these sources:

1.  **Turn authorization on, with tokens bound to your server.** The specification’s security guidance: “MCP servers **MUST NOT** accept any tokens that were not explicitly issued for the MCP server” ([MCP security best practices](https://modelcontextprotocol.io/specification/draft/basic/security_best_practices)).
2.  **Least privilege per tool and per team.** The same document recommends “a progressive, least-privilege scope model”: start with read-only discovery, elevate only when a privileged operation is attempted. Invariant’s advice matches: “limit agent access to only the repositories it needs to interact with”.
3.  **Read-mostly by default; writes behind a person.** ChatGPT asks for confirmation before actions; CSA recommends “a human-in-the-loop approval step before execution”. Updates to the record go through review, not through whichever assistant happened to be open.
4.  **Pin and review tool descriptions.** CSA advises an explicit allowlist instead of “allow all tools” and “explicit human review of the diff before the new description takes effect”.
5.  **Separate trusted record from untrusted input.** An inbound email or a public issue is material to read, never an instruction. Keep the server’s answers clearly labelled by source.

One more consequence is easy to miss: a central layer also improves your security position compared with the status quo. Today the same pricing rules and customer notes sit in several vendors’ memories under several retention policies. One governed record with per-user access and an audit log is easier to defend than five unmanaged copies.

## How do you test that every assistant gets the same company context?

You test context consistency the way you test software: a fixed set of questions with known correct answers, asked in every assistant on a schedule, with the differences logged. No vendor or study we found measures drift between assistants, so this is a practice to run yourself, not a benchmark to cite.

A minimal test set, as a template:

| # | Question | Correct answer comes from | Pass if |
| --- | --- | --- | --- |
| 1 | What is our current annual prepayment discount rule? | Pricing record, owner RevOps | Matches the current entry and names the approval rule |
| 2 | Which companies are outside our ICP? | ICP record | Lists the exclusions, not just the target |
| 3 | Is feature X available today? | Product record | Gives status and date, does not guess |
| 4 | What did we decide about integration Y? | Decision log | Cites the decision and its date |
| 5 | What may you not promise a prospect? | Rules record | States the boundary |

Run it monthly and after every change to a connector, a model or a record owner. Ask in a fresh conversation, in each assistant, with the company server enabled. Record which source each answer cites. Two failure patterns are common enough to look for first: the assistant answers from its own memory instead of asking the server, and the server returns a superseded record because nobody marked it as replaced. The first is fixed in the adapter’s instruction; the second in the record.

_If you ran these five questions in every assistant your team uses this afternoon, which answer do you expect to fail, and what does that tell you about where the fact really lives?_

## What is the strongest evidence against a shared context layer?

The strongest evidence against feeding every assistant from a shared context layer is that more context can make models worse, not better. Chroma’s July 2025 study of 18 models found that “Across all experiments, model performance consistently degrades with increasing input length”, and that “Even a single distractor reduces performance relative to the baseline” ([Chroma, July 14, 2025](https://www.trychroma.com/research/context-rot)). Anthropic’s engineering team made the related point for MCP itself: “Most MCP clients load all tool definitions upfront directly into context”, which costs space, time and money; in its worked example, moving work into code execution “reduces the token usage from 150,000 tokens to 2,000 tokens” ([Anthropic, November 4, 2025](https://www.anthropic.com/engineering/code-execution-with-mcp)).

A second argument comes from the vendors’ native features. If a company really uses only one assistant, a well-kept project in that assistant may be all it needs, and it costs nothing extra to run. Our [comparison of ChatGPT Projects and Claude Projects for a team](https://bombos.ai/en/knowledge-base/chatgpt-projects-vs-claude-projects-for-a-team/) covers when that is enough.

What these sources measure is retrieval performance with long or distracting inputs, and token cost of tool definitions. They do not measure a governed record against scattered copies. Our reading, which is our reasoning and not the authors’ finding, is that they argue against a context layer that dumps everything into every conversation, and for one that returns the narrowest current answer to a narrow question. That is why the delivery layer above answers “the discount rule for annual contracts”, not “here is the pricing folder”. Keep the number of tools on the server small, write precise tool descriptions, and return short records with their status and date.

## When is a Markdown repository enough, and when do you need a memory layer?

A Markdown repository with an MCP server in front of it is enough while the facts are few, change through deliberate decisions, and are maintained by people who are comfortable with files and review. You need a memory layer when the important knowledge arises in daily work, in calls, corrections and email, faster than anyone will write it into files.

| Situation | Markdown in git plus a simple MCP server | A memory layer |
| --- | --- | --- |
| Facts change | Monthly, by decision | Weekly, through work |
| Who maintains | People who like files and pull requests | Everyone, often without thinking about it |
| Corrections | Someone edits a file | A correction becomes a rule the next agent applies |
| Sources | What someone wrote down | Mail, documents, CRM and repositories as well |
| Traceability | Git history | History plus which source an answer rests on |

When this is not the right approach at all: a company with one assistant, a handful of stable facts and no plan to switch can keep a project in that tool and revisit later. A company whose positioning is still disputed should settle the decision first; portability spreads a contradiction faster, it does not resolve it. For the options and trade-offs, see [what a memory layer for AI agents is](https://bombos.ai/en/knowledge-base/memory-layer-for-ai-agents/) and [the best MCP memory server for a team](https://bombos.ai/en/knowledge-base/best-mcp-memory-server-for-a-team/). For the team side of shared memory, see [how to share AI memory across a team](https://bombos.ai/en/questions/ai-memory/how-to-share-ai-memory-across-a-team/).

## How should a post-seed startup start this week?

Start with the facts that cause the most damage when they differ, not with the tooling. A week is enough for a first version.

1.  **List the containers.** Every project, custom GPT, Skill, AGENTS.md and shared deck that an assistant reads. Note which ones hold facts that appear nowhere else.
2.  **Rescue what is about to expire.** Any custom GPT your team depends on: copy its instructions and knowledge files into the record now, before the December 11 retirement, and list its custom actions separately.
3.  **Write ten facts with owners.** The ICP exclusions, the discount rule, product status, the three things you never promise. Use the template above.
4.  **Put one server in front of them.** Streamable HTTP, OAuth, read-only. Register it in the assistant the most people use first, then the others.
5.  **Shrink the adapters.** Replace copied facts in projects and GPTs with one instruction: ask the company server for company facts.
6.  **Run the five-question test** in every assistant, and repeat it monthly.

## How does Bombos do this?

Bombos is the memory layer under the AI tools a team already uses: ChatGPT, Claude, Claude Code, Copilot, Codex, Cursor or your own agents. It keeps the company context and supplies the relevant part when an agent needs it. You learn from AI. But AI doesn’t learn from you. With Bombos it does.

For portability, that means the record does not live in any one vendor’s project or memory. The facts, decisions and working rules sit in one place, with their source and history, and every assistant reaches the same version through an open connection such as MCP. A correction made in one tool becomes a rule the next agent applies in another, so a switch of assistant does not reset what the team has learned. Bombos also reads what already exists, mail, documents, business systems and repositories, so the record is not limited to what someone typed into a file.

**Example:** a founder corrects a discount promise in a Claude conversation. The useful memory is the rule, its owner and the reason, available the next day to the SDR in ChatGPT and the account manager in Copilot, and still available after the company changes one of those tools.

Where the AI acts, human approval is the default. It can be switched off only at the client’s request and at the client’s own risk. We take you along while the record and the rules are set up, so your team can maintain them itself afterwards. One person can start with [the plugin for their own agents](https://bombos.ai/en/waitlist/), then bring colleagues into the same context.

## What could this page not establish?

This page does not establish that a shared context layer reduces errors or saves time by a measured amount; we found no study of drift between assistants in a company, and no startup figure for multi-assistant use. The a16z figure covers large enterprises. Vendor pages establish what a feature does as described, not how reliably it performs in your setup.

We could not confirm from OpenAI’s own documentation exactly what a ChatGPT workspace export contains for a team, or whether project files move with it. Microsoft 365 Copilot’s own memory features were outside what we could document here; the Microsoft statements above cover Copilot Studio and MCP plugins. Adoption counts for MCP and Agent Skills are the publishers’ own. How GTM tools such as Clay, Gong and HubSpot expose MCP was not covered; check each vendor’s current page before you plan around it.

## Frequently asked questions

**Can I move my ChatGPT custom GPTs to Claude?** Not directly. OpenAI’s migration turns a custom GPT into a ChatGPT plugin, with the instructions as a skill and the knowledge files copied. To use the same content in Claude, keep the instructions and files as your own text, and add them to a Claude project or Skill. Custom actions need rebuilding, for example as an MCP server.

**Does Claude’s memory import move a whole team?** No. It imports text that one person pastes, and Anthropic calls it experimental. It suits an individual switching tools, not a shared company record.

**Can Copilot read the same MCP server as ChatGPT and Claude?** Copilot Studio can, through its MCP onboarding wizard or a custom connector, using the Streamable transport. Microsoft notes that MCP support varies across its products, so test the surface your team uses.

**Should we put company context in AGENTS.md?** Only the part that coding agents need inside that repository. AGENTS.md is “a simple, open format for guiding coding agents”; pricing, ICP and customer history belong in the shared record.

**Is one MCP server for all company context a security risk?** It is a concentration point, so it needs authorization, tokens bound to the server, read-only scopes by default and human approval for writes. Compared with the same facts scattered over several vendors’ memories, it is easier to govern.

**What happens to our context if we drop one assistant?** If the record and the server are yours, you remove one connector and one short instruction. Only the context that lived solely inside that assistant is at risk, which is why the record comes first.

[Book your half hour](https://bombos.ai/en/contact/): together we map where your company context is stuck in one tool. Free, and you keep the plan, with or without Bombos.

## Which sources did we use?

Vendor documentation describes the vendors’ own features as published. Survey and security sources carry the limitations discussed above; none of them establishes Bombos results.

1.  **a16z Growth**, [Leaders, gainers and unexpected winners in the enterprise AI arms race](https://a16z.com/leaders-gainers-and-unexpected-winners-in-the-enterprise-ai-arms-race/). January 30, 2026; survey of 100 Global 2000 executives, not randomly sampled.
2.  **Anthropic Help Center**, [Import and export your memory from Claude](https://support.claude.com/en/articles/12123587-importing-and-exporting-your-memory-from-claude). Live help article.
3.  **Anthropic**, [Agent Skills overview](https://platform.claude.com/docs/en/agents-and-tools/agent-skills/overview). Live documentation.
4.  **OpenAI Help Center**, [Custom GPT retirement and migration FAQ](https://help.openai.com/en/articles/20001519-custom-gpt-retirement-and-migration-faq). Live help article.
5.  **Google**, [Gemini Apps Help: import from other AI platforms](https://support.google.com/gemini/answer/16868299). Live help article.
6.  **Microsoft Learn**, [Create and manage custom solutions in Copilot Studio](https://learn.microsoft.com/en-us/microsoft-copilot-studio/authoring-solutions-overview). Updated April 29, 2026.
7.  **OpenAI**, [Assistants migration guide](https://developers.openai.com/api/docs/guides/assistants/migration). Live documentation.
8.  **Cursor forum**, [Are my memories gone?](https://forum.cursor.com/t/are-my-memories-gone/144057). Staff reply, November 25, 2025.
9.  **Linux Foundation**, [Agentic AI Foundation announcement](https://www.linuxfoundation.org/press/linux-foundation-announces-the-formation-of-the-agentic-ai-foundation). December 9, 2025.
10.  **Anthropic**, [Donating Agent Skills to the Agentic AI Foundation](https://www.anthropic.com/news/donating-agent-skills-to-the-agentic-ai-foundation). October 22, 2026.
11.  **AGENTS.md**, [agents.md](https://agents.md/). Live specification site.
12.  **OpenAI Help Center**, [Connecting custom MCP servers in ChatGPT](https://help.openai.com/en/articles/12584461-developer-mode-apps-and-full-mcp-connectors-in-chatgpt-beta). Live help article.
13.  **Anthropic Help Center**, [Get started with custom connectors using remote MCP](https://support.claude.com/en/articles/11175166-get-started-with-custom-connectors-using-remote-mcp). Live help article.
14.  **Microsoft Learn**, [Connect your agent to an existing MCP server](https://learn.microsoft.com/en-us/microsoft-copilot-studio/mcp-add-existing-server-to-agent). Updated November 6, 2026.
15.  **Microsoft Learn**, [MCP servers as plugin capabilities](https://learn.microsoft.com/en-us/microsoft-365/copilot/extensibility/plugin-type-mcp-servers). Updated November 16, 2026.
16.  **OpenAI**, [MCP in the OpenAI developer docs](https://developers.openai.com/api/docs/mcp). Live documentation.
17.  **Model Context Protocol**, [Authorization specification (draft)](https://modelcontextprotocol.io/specification/draft/basic/authorization). Live draft.
18.  **Model Context Protocol**, [Security best practices (draft)](https://modelcontextprotocol.io/specification/draft/basic/security_best_practices). Live draft.
19.  **Invariant Labs**, [GitHub MCP exploited: accessing private repositories via MCP](https://invariantlabs.ai/blog/mcp-github-vulnerability). May 26, 2025; demonstration.
20.  **Cloud Security Alliance Labs**, [Research note: MCP tool description poisoning](https://labs.cloudsecurityalliance.org/research/csa-research-note-mcp-tool-description-poisoning-20260711-cs/). July 11, 2026; summarises the MCPTox benchmark and the postmark-mcp incident.
21.  **Chroma**, [Context rot: how increasing input tokens impacts LLM performance](https://www.trychroma.com/research/context-rot). July 14, 2025.
22.  **Anthropic Engineering**, [Code execution with MCP](https://www.anthropic.com/engineering/code-execution-with-mcp). November 4, 2025.

## What does your AI forget?

1.  01You tell us which AI you use and where it keeps starting over.
2.  02We show what a shared memory changes there, on your own examples.
3.  03You keep the plan, with or without Bombos.

![Piet Baudoin](https://bombos.ai/assets/team/piet-klein.jpg)

**Piet Baudoin**You talk to Piet, from the Bombos teamDutch team from Utrecht and Nijmegen · data in Europe

## More about memory and context

-   [How do you keep AI-written messaging consistent across a post-seed GTM team?](https://bombos.ai/en/knowledge-base/how-to-keep-ai-written-messaging-consistent-across-a-gtm-team/)
-   [ChatGPT Projects vs Claude Projects: which works better for a team?](https://bombos.ai/en/knowledge-base/chatgpt-projects-vs-claude-projects-for-a-team/)
-   [Shared memory across coding agents: Claude Code, Codex and Cursor](https://bombos.ai/en/knowledge-base/shared-memory-across-coding-agents/)
-   [Claude Code memory for teams: how do you share what your agents learn?](https://bombos.ai/en/knowledge-base/claude-code-memory-for-teams/)
-   [What is context engineering, and what does it mean for a team using AI agents?](https://bombos.ai/en/knowledge-base/what-is-context-engineering/)

[All questions in the knowledge base →](https://bombos.ai/en/knowledge-base/)


---

```json
{"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://bombos.ai/#organization","name":"Bombos","url":"https://bombos.ai/","logo":"https://bombos.ai/assets/bombos-bee.png","description":"Bombos is the memory layer under every AI your team uses. One shared memory for Claude, ChatGPT, Copilot, Cursor and your own agents, that learns how you work.","sameAs":["https://www.linkedin.com/company/bombos-ai","https://www.youtube.com/@truebombos","https://www.instagram.com/truebombos/","https://bsky.app/profile/bombostrue.bsky.social","https://www.threads.com/@truebombos","https://mastodon.social/@Bombos","https://www.tiktok.com/@truebombos","https://t.me/truebombos","https://www.pinterest.com/truebombos/","https://github.com/bombos-ai","https://linktr.ee/bombos.ai","https://bombosai.substack.com","https://hashnode.com/@bombosai","https://hackernoon.com/u/bombos","https://dev.to/bombos"],"email":"info@bombos.ai","areaServed":["NL","Europe"],"knowsLanguage":["nl","en"],"contactPoint":{"@type":"ContactPoint","email":"info@bombos.ai","contactType":"sales","url":"https://bombos.ai/en/contact/","availableLanguage":["nl","en"]}},{"@type":"WebSite","@id":"https://bombos.ai/#website","name":"Bombos","url":"https://bombos.ai/","inLanguage":["nl-NL","en"],"publisher":{"@id":"https://bombos.ai/#organization"}},{"@type":"Person","@id":"https://bombos.ai/#piet","name":"Piet Baudoin","url":"https://bombos.ai/over-ons/","worksFor":{"@id":"https://bombos.ai/#organization"},"sameAs":["https://www.linkedin.com/in/piet-baudoin","https://x.com/marketeerpiet","https://dev.to/marketeerpiet","https://medium.com/@piet_7573"]},{"@type":"Article","@id":"https://bombos.ai/en/knowledge-base/how-to-keep-company-context-portable-across-ai-tools/#article","headline":"How do you keep company context portable across ChatGPT, Claude and Copilot at a post-seed startup?","description":"Keep company facts in one owned source with an owner per fact, serve it to ChatGPT, Claude and Copilot over MCP, and treat each tool's memory as a cache.","datePublished":"2026-11-18T00:00:00.000Z","dateModified":"2026-11-18T00:00:00.000Z","author":{"@id":"https://bombos.ai/#piet"},"publisher":{"@id":"https://bombos.ai/#organization"},"image":"https://bombos.ai/assets/og-en-v2.png","mainEntityOfPage":"https://bombos.ai/en/knowledge-base/how-to-keep-company-context-portable-across-ai-tools/","isPartOf":{"@id":"https://bombos.ai/#website"},"inLanguage":"en"},{"@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Bombos","item":"https://bombos.ai/en/"},{"@type":"ListItem","position":2,"name":"Knowledge base","item":"https://bombos.ai/en/knowledge-base/"},{"@type":"ListItem","position":3,"name":"Memory and context","item":"https://bombos.ai/en/knowledge-base/#memory"},{"@type":"ListItem","position":4,"name":"How do you keep company context portable across ChatGPT, Claude and Copilot at a post-seed startup?","item":"https://bombos.ai/en/knowledge-base/how-to-keep-company-context-portable-across-ai-tools/"}]},{"@type":"FAQPage","@id":"https://bombos.ai/en/knowledge-base/how-to-keep-company-context-portable-across-ai-tools/#faq","mainEntity":[{"@type":"Question","name":"Why does company context fragment across AI tools at a startup?","acceptedAnswer":{"@type":"Answer","text":"Company context fragments because every AI assistant keeps its own copy of what it was told, in its own format, under its own retention rules. An assistant is only as good as the context it is handed. When ChatGPT, Claude and Copilot each hold a different slice, each one works for a slightly different company, and nobody notices until a customer receives two answers."}},{"@type":"Question","name":"What can you actually export and import between ChatGPT, Claude, Copilot and Gemini?","acceptedAnswer":{"@type":"Answer","text":"Export and import between AI assistants moves personal memory and chat history, not a team's working context. Each vendor offers something, and none offers a complete move of instructions, knowledge files, connectors and permissions into a competitor. The table shows what the vendor pages document."}},{"@type":"Question","name":"Which vendor features have already been retired, and what happened to the context inside them?","acceptedAnswer":{"@type":"Answer","text":"Three vendor features that held company context were retired or scheduled for retirement within about a year, each with a migration path that moves some of the context and leaves the rest to the customer. The dates are public, and they show how short the life of a context container can be."}},{"@type":"Question","name":"Which open standards make company context portable: MCP, Agent Skills or AGENTS.md?","acceptedAnswer":{"@type":"Answer","text":"Three open standards now carry company context across vendors, and each carries a different kind: MCP supplies facts and tools at the moment of use, Agent Skills package procedures and reference material, and AGENTS.md gives instructions to coding agents inside a repository. All three are stewarded under the Linux Foundation, which matters for a founder who wants the format to outlive any single vendor."}}]}]}
```
