Free Intensive Nov 21–22Free 2-day hands-on Vibecoding Intensive on Nov 21–22 at 16:00 MSK (Sat–Sun) · Live Cohort 3 starts Nov 23Join Free →Cohort 3 (Nov 23)

How to Connect Any AI Agent to Your Own Memory Through One MCP with Two Tools

Memory and skills inside ChatGPT or Claude are convenient, but they tie you to a single service. The author of the "LLM under the hood" channel rents only the agent's "core" and keeps the context on his own side, serving it through one MCP (the standard connector between AI and tools) with two tools. This MCP acts as a loader: Claude on a phone or Codex on a laptop pull in the environment as needed. The core can be swapped, and the context stays yours.

AI-processed from TG: @llm_under_hood (LLM под капотом); edited by Hamidun News
How to Connect Any AI Agent to Your Own Memory Through One MCP with Two Tools
Source: TG: @llm_under_hood (LLM под капотом). Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceTG: @llm_under_hood (LLM под капотом) ↗
Editor-in-Chief
Corethe Agents SDK from OpenAI, which powers ChatGPT and Codex
Key metric #2The loader grants different permissions and tools depending on who is launching it and from where
Multiple cores sharing one contexta swarm, with processes, which he calls Company Brain

Imagine you spent six months setting up memory, skills, and connected services in ChatGPT, and then a different model became the best one. Switching becomes a project, because everything you've accumulated lives inside a specific product. The author of the Telegram channel "LLM under the hood" ran into this while building a comfortable way of working with agents, first for himself and then for teams. His solution fits into one sentence: rent the agent core from the best vendors on the market, and use your own context, memory, and tools. To do this, he wires one MCP server (MCP is the standard protocol through which AI applications connect to external tools and data) with two tools into every client.

How the usual setup works

In the standard picture, the agent sits at the center. Most often this is ChatGPT or Claude with shared memory and tools that are always at hand, and that's how most people work. Less commonly, there is a custom agent harness (the software wrapper around the model) on a server: Codex, Claude Code, Hermes, and similar tools. You connect to it through a terminal, a chat, or a web interface. There are also cloud agents like Dot, something between the first two options, and very rare homegrown solutions.

The main convenience of systems centered on an agent is that integrations and tools work out of the box, and few people think about how it all works. But look closer and all these products share a common core. According to the author, ChatGPT, Codex, and the others run on the same foundation: the Agents SDK from OpenAI. Memory and skills attach to the core from the outside. OpenAI's own staff, as the post notes, say models are improving so fast that complex SDKs and harnesses may soon become less necessary. If that holds, the value shifts from the wrapper to what you load into it, that is, to your context.

One MCP instead of built-in memory

The author rents the core through a subscription (Claude and ChatGPT), but he doesn't use the built-in memory, skills, and tools of these services. What matters to him is being able to plug in any core and get access to his context whether from Claude on mobile, from Dot, or from Codex on a laptop.

To achieve this, each client gets one MCP server with just two tools. It is a kind of "loader," lightly optimized for the agent core. Through it, and following the principle of progressive disclosure, the core loads the entire working environment and starts working with it. A everyday analogy: a new employee isn't handed the company's entire archive on day one. They get a badge and are shown where things are, and the rest they open as needed.

There is also a bonus for teams. Because access goes through a single entry point, the loader can define different permissions and different tools depending on who is launching it and from where. The same context is available from a mobile client and from a laptop, but the capabilities in each case may differ.

From a swarm to Company Brain

If you run several different cores over a shared context, you get a swarm of agents: the author has already covered such experiments in his channel. And if you describe the workflows in detail, along with the rules for handing tasks between people and agents, you arrive at what Y Combinator calls a Company Brain.

What to take away from this

For a developer or product manager, the takeaway is practical. Before embedding a specific provider's memory and skills into your processes, ask yourself whether you could take them with you. If the context lives in your own infrastructure and is accessible through a standard protocol like MCP, switching models becomes a matter of switching subscriptions.

There are also honest limitations. This is a description of the author's personal practice, not a benchmark: the post contains no figures on token savings, speed, or cost. You will have to build and maintain your own loader, access controls, and context storage, while ChatGPT's and Claude's built-in features come free and ready to use. The approach is aimed at people willing to own their own memory layer and answer for its quality. But for teams that don't want to depend on a single vendor, the idea of "rent the core, own the context" looks like a sensible starting point.

Story

Key Questions Answered

How the usual setup works?

In the standard picture, the agent sits at the center. Most often this is ChatGPT or Claude with shared memory and tools that are always at hand, and that's how most people work. Less commonly, there is a custom agent harness (the software wrapper around the model) on a server: Codex, Claude Code, Hermes, and similar tools. You connect to it through a terminal, a chat, or a web interface. There are also cloud agents like Dot, something between the first two options, and very rare homegrown solutions.

What are the key details on One MCP instead of built-in memory?

The author rents the core through a subscription (Claude and ChatGPT), but he doesn't use the built-in memory, skills, and tools of these services. What matters to him is being able to plug in any core and get access to his context whether from Claude on mobile, from Dot, or from Codex on a laptop. To achieve this, each client gets one MCP server with just two tools. It is a kind of "loader," lightly optimized for the agent core.

What are the key details on From a swarm to Company Brain?

If you run several different cores over a shared context, you get a swarm of agents: the author has already covered such experiments in his channel. And if you describe the workflows in detail, along with the rules for handing tasks between people and agents, you arrive at what Y Combinator calls a Company Brain.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Hands-On AI Engineering · Hamidun Academy

Build & Ship Production AI Agents Without Boilerplate Code

Start your free 72-hour trial at Hamidun Academy (academy.hamidun.com) — interactive lessons on Claude Code & Cursor, live Tech Radar, and production agent templates.

Start 72h Free Trial →Vibe-Coding Intensive
What do you think?
Loading comments…