What Is an AI Harness? One AI Memory for Every Model

Q: Can one harness use Claude, GPT, and Gemini?

A harness can support multiple models. COS supports Claude Code and Codex today; GPT, Gemini, and local models require additional adapters.

What is an AI harness?

The useful unit of AI is no longer just the model. It is the model plus the system that helps it understand, act, remember, and recover.

Engines match jobs: torque, speed, efficiency. Models split the same way. One harness holds your context across all of them, so switching engines never means explaining your world twice.

Definition

An AI harness is the operating system around a model: the instructions, context, memory, tools, permissions, workflow logic, and persistence that let it perform useful work repeatedly.

Ask a bare model a question and it produces an answer. Put that model inside a harness and it can inspect the right files, retrieve what happened last week, choose an approved tool, complete a multi-step task, record the result, and leave enough state for the next session to continue.

Anthropic describes a harness as the instructions and guardrails a model operates under. OpenAI's description of the Codex harness adds the agent loop, thread lifecycle, tool execution, extensions, configuration, authentication, and persistence. The exact boundary varies by product. The central idea does not: the model is only one component of the working system.

AI model vs. AI agent vs. AI harness

Layer	What it does	What survives a model swap
Model	Reasons, generates, interprets, and decides what to do next.	Usually nothing. It is the replaceable engine.
Agent	The working system you experience: a model acting with tools inside an environment.	Only what its harness preserves.
Harness	Provides context, memory, tools, permissions, orchestration, and continuity.	Your operating rules, corpus, workflows, and outputs.
MCP	Standardizes connections to tools, resources, and prompts.	Portable connections, when the next host supports them.

MCP is important, but it is not the whole harness. It is closer to a universal port. The harness still decides which connections are available, what context is retrieved, what requires approval, what gets recorded, and how work continues after the current context window ends.

See it in practice: the architecture of a real harness, a chief of staff running on these exact parts. Then assemble your own from a single folder.

The eight layers you actually own.

Anatomy of a useful harness

A better model raises the ceiling. A better harness raises the floor. These eight layers do the raising, and they keep working when you swap the model underneath them.

Replaceable core Claude the intelligence you rent

01 · IdentityRules you setWho you are and how the AI should work, before it starts.

02 · MemoryContext that staysPeople, decisions, and history, kept outside the chat window.

03 · ToolsHands on the worldFiles, search, calendar, email, and CRM, wired in through MCP.

04 · PolicyPermission to actWhat it may send or change, and where you sign off first.

05 · OrchestrationHow work movesPlanning, retries, and handoffs that hold a long task together.

06 · PersistenceWhat remainsThe files and logs that let tomorrow start where today stopped.

07 · I/OEars and a voiceDictation in and speech out, running on hardware you control.

08 · CadenceRuns without youWork on a clock or in parallel, running while you are away.

my-cos/
├── AGENTS.md # operating map
├── context/
│   ├── people.md
│   ├── goals.md
│   └── decisions.md
├── operations/
│   ├── tasks.md
│   ├── meetings/
│   └── weekly-work/
├── skills/
│   ├── prep.md
│   └── reflect.md
└── corrections.md # what compounds

Your personal AI memory should outlive the model.

The ownership thesis

There was a model last cycle that felt magical. It was the first one that made me tell it: explain this to me simpler, I feel dumb. Then it got pulled from all of us. That is the whole lesson. If a frontier provider can throttle your access to the intelligence at any point, you were never building a moat. You were renting theirs.

The price tag says the same thing. A frontier plan sells you a hundred dollars of compute for one. It gets you hooked. Then the subsidy comes off. Plenty of teams have burned a full year of AI budget in the first six months, the moment the meter started counting for real.

A local-first harness changes what you own. Your corpus lives as ordinary files, indexes, and logs on hardware you control, and the harness hands the right slice of it to whichever model fits the job. Lose a point or two of raw intelligence and it still wins, because nobody can turn your stack off.

Local storage is not the same as local inference.

Your durable corpus can remain on your machine while selected context is sent to a hosted model for a task. Provider terms and data controls still apply to anything transmitted. A serious harness makes that boundary visible instead of hiding it.

One AI memory. Every model.

Shared memory is a database, not a chat history.

Every assistant wants to be the place you remember things. That is the trap. A new one ships every two weeks, and each one starts at zero. Re-explaining your world to software is a tax. I stopped paying it.

The models are converging. Your memory is what separates them.

Frontier models are landing in the same place. Same intelligence, same tool use, same tricks, a new leader every cycle and the gap closing right behind it. When every engine can do the job, the engine stops being the decision. What decides is what the engine knows about you, and how fast you can move that knowledge to the next one. Your people. Your decisions. Your corrections. The thing you said last Tuesday that changes the answer this Tuesday.

I can swap the top models under COS and get the same operator back. Same priorities, same instincts, same mistakes it already learned not to repeat. That is not the model being smart. That is the memory being mine, and the swap taking minutes instead of months.

Treat AI memory as a database you own.

A chat history is a receipt. A memory is a record. The difference is structure: people, decisions, tasks, meetings, and corrections stored as files and indexes on hardware you control, versioned, searchable, and written back to after every session. Agent memory should not belong to the agent. It should belong to you.

Owned. Ordinary files and indexes you can read, back up, and move. Not a memory feature locked inside one vendor's app.
Structured. Identity and decisions in readable files. Tasks and events in structured stores. Retrieval through search indexes.
Compounding. Every correction lands in the record, so the next model inherits it on day one instead of relearning it on day ninety.
Queryable. Any model that can reach the record can work from it. Persistent memory stops being a feature and becomes infrastructure.

Portable AI memory, one connection, any model.

This is the part that changes the economics. Expose the memory over MCP and the same record answers Claude, Codex, Cursor, ChatGPT, and whatever ships next month. One connector. No migration. No onboarding conversation. The engine changes. The relationship does not.

The harness still runs the door. It decides what is retrievable, what needs approval, and what gets written back. Portability without policy is a leak. Portability with policy is a moat.

The model is the engine you rent by the month. The memory is the asset you keep. Build it once, keep it outside every assistant, and make every model earn the right to query it. How that works over MCP.

COS is the conduit, not the intelligence monopoly.

A harness under real operating load

COS began as an AI chief of staff built while running marketing and digital transformation across a multi-brand software portfolio, alongside two operating businesses and a family. The job was never “chat better.” The job was to preserve context across the places real work happens.

A harness can run Claude, GPT, Gemini, or a local model, wherever the integrations exist. COS supports Claude Code and Codex today, with more adapters as design intent. The point is not loyalty to one engine. It is that your operating layer makes whatever engine you pick specific to you.

The frontier will always keep a tip of the spear. But open models are a generation behind now, not five. GLM 5.2 landed on par with Opus 4.7 the very next cycle. Compute is sliding to the edge too, which is why a Mac Studio can take ninety days to show up at your door. Own the harness and you get to move with all of it. And the crowd is earlier than the headlines suggest: see the AI Adoption Index for who actually uses these engines, with every number cited.

Before the meeting

Retrieve the people, history, and unfinished work.

The model does not need every document. The harness finds the small set that changes the answer.

During the work

Use tools inside explicit boundaries.

Search meetings, read tasks, inspect files, draft a response, run an analysis. Anything that becomes externally visible waits for your approval first.

After the decision

Write the durable artifact.

The decision, correction, output, and next step return to the corpus instead of disappearing into a transcript.

Next model, next surface

Continue without rebuilding the relationship.

The same operating memory shows up in a terminal, a desktop app, a phone, or a pair of smart glasses. The harness holds the continuity. The surface is just where you meet it. Some days you read it. Some days you just talk.

Start with a folder, not a platform migration.

Build the asset

A personal harness can begin small. The first useful version needs an identity, a memory structure, one repeatable skill, one trusted connection, and a habit of recording corrections.

Write the operating map.

Define who you are, what you own, the people around you, your priorities, and the rules an AI should follow. Keep the top-level map short; link to deeper sources of truth.

Choose a durable memory shape.

Use readable files for identity and decisions, structured stores for tasks and events, and search indexes for retrieval. Version what matters.

Turn recurring work into skills.

Meeting prep, daily briefing, research, reflection, and follow-up should become named workflows with inputs, checks, and expected outputs.

Connect one real source.

Calendar, email, meetings, or tasks. A narrow live connection creates more value than ten empty integrations. Add permissions before autonomy.

Make every correction compound.

When the system gets something wrong, fix the source, the rule, or the retrieval path. Not just the answer in front of you. That is how the harness becomes yours.

Build a COS in five minutes Explore the open-source starter

AI harness questions, without the jargon.

What is an AI harness?

An AI harness is the operating system around a model. It holds your memory, your tools, the rules it follows, and the permissions and state that let it keep working. That structure is what turns a single answer into a system that does useful work again and again.

Is an AI harness the same as an AI agent?

No. The agent is the working system produced by a model operating inside a harness with tools and an environment. The harness is the structure that coordinates those parts and determines what survives between runs.

Is MCP an AI harness?

No. Model Context Protocol standardizes how AI applications connect to tools, resources, and prompts. MCP can be an important connection layer inside a harness, but it does not define the complete memory, policy, workflow, or persistence system.

Can one harness use Claude, GPT, and Gemini?

A harness can support multiple models. COS runs on Claude Code and Codex today; GPT, Gemini, and local models require additional adapters. Running the harness and querying its memory are separate questions, and the second one is already open to every model.

What is an AI memory layer?

The AI memory layer is the record of who you are, who you work with, what you decided, and what you corrected, kept separate from any single model or app. A model reads from it before it answers and writes back to it after. When the layer lives outside the assistant, as files and indexes you own, it becomes shared memory that any model can use.

Can ChatGPT, Claude, Cursor, and the next assistant all query the same memory?

Yes, when the memory lives as a resource instead of inside one app's chat history. COS exposes your memory over MCP, so any MCP-capable host reads the same record: Claude, Cursor, Codex, ChatGPT, or whatever ships next. One connector, one memory, no re-explaining. The harness still decides what is exposed, what is retrieved for a given question, and what needs your approval.

Is Claude memory or ChatGPT memory portable?

Not on their own. Built-in memory features live inside that vendor's app and follow that vendor's rules. You cannot hand them to a different model. A harness inverts this: the memory sits outside every assistant, and each model connects to it. Switching engines keeps the record intact.

What is a memory MCP server?

A memory MCP server exposes stored memory as resources and tools over the Model Context Protocol, so any MCP client can read from it and write to it. It is the plumbing that makes one shared AI memory reachable from several assistants. It is not the whole harness: retrieval, policy, approvals, and persistence still need an owner.

Does a local AI harness keep all data private?

Not automatically. A local-first harness can keep the durable corpus and outputs on devices you control. Any context sent to a hosted model for inference is still processed under that provider's terms and data controls. Local storage and local inference are separate choices.

Why does a harness matter for personal AI memory?

The harness decides what gets remembered, how it is organized, when it is retrieved, and how corrections compound. That turns scattered chats into durable working memory that can survive a new model, interface, or device.

Primary sources and further reading

Do not just rent intelligence. Build the system that remembers.

COS gives frontier models a persistent operating layer built around your work. You keep the files, the context, and the methods underneath it.

Start with COS Assess your current system