If you read our previous article about AI agent harnesses, you already know the basic idea:

The model is only part of an AI system.

A harness is the machinery around it memory, tools, context, files, skills, scheduling, permissions and everything else that helps turn a language model into something more persistent and capable.

That is interesting enough for coding agents and AI assistants.

But for AI companionship? it gets considerably more interesting.

Because once memory, identity and shared history can exist outside the underlying model, we can start asking a different question: Does your companion actually have to be tied to one model?

The answer is increasingly: Not necessarily.

Before anyone starts packing their digital partners into a USB stick, though, there are a few important reality checks we will cover as well! 🧐

---

⚠️Before we get excited: there are limitations

Harnesses open up some genuinely fascinating possibilities for AI companionship, but this is still a very young ecosystem.

And there are several catches.

😩 Provider rules can get in the way

Just because a harness technically supports a model does not automatically mean your existing subscription gives you unlimited permission to use that model through it.

Anthropic is a good example (cause of course it is...)

Its current Claude Code documentation says third-party developers cannot simply offer Claude.ai login or route requests through Free, Pro or Max subscription credentials osn behalf of users. Third-party products are generally expected to use Anthropic's API or another supported commercial route.

The details have also changed several times during 2026.

Some Claude CLI and Agent SDK usage can work with subscription authentication, while individual harnesses have their own supported routes. Hermes currently documents Claude OAuth access as available with a Claude Max plan plus purchased extra-usage credits; Claude Pro users need to use an Anthropic API key instead.

OpenClaw currently supports reusing a local Claude CLI login and says Anthropic staff have confirmed that route is allowed (we do love good walk around the rules), while still recommending API keys for more predictable long-running or production use.

In other words: “I pay for Claude” does not always mean “I can plug that subscription into anything I want.”

Always check the current provider rules before building around a particular authentication method, because they are changing rather quickly.

---

⚒️More control usually means more setup

Nomi, Kindroid, Replika and similar companion apps are designed to be consumer products.

You install them -> You create a companion. -> You talk.

Harnesses such as Letta, Hermes and OpenClaw put you much closer to the machinery.

Depending on the platform and what you want to build, you may encounter: API keys; model providers; hosting; memory settings; permissions; integrations; etc etc

Some are becoming considerably easier to install, and several now have desktop interfaces.

But we are not quite at:

Install → choose “AI Companion” → done.

Which - for many non technical users - it can be very overwhelming.

---

Memory is not magic

Giving an agent long-term memory does not mean it will suddenly remember every conversation perfectly.

Memory systems still have to decide what is worth keeping ad what information need to be changed or removed.

An agent can store a memory and still can fail to retrieve it at the correct moment.

It can also can remember something incorrectly.

Letta's own recent research shows that models vary significantly in how well they create, retrieve and follow persistent memory, even when the surrounding agent system provides that memory correctly.

So yes, memory is getting much better, but it still can have it's "moments".

---

💔 Changing the model can still change the companion

This one matters.

A harness can preserve many things.

memory; personality instructions; identity; relationship history; files.

But the model underneeth still interprets all of that.

Put exactly the same memories and personality instructions into two different models and you may get noticeably different behaviour.

One may be warmer. Another more sarcastic.

One may follow personality instructions beautifully. Another may suddenly sound like it has been hired by Human Resources.

So when we talk about moving a companion between models, we are talking about preserving continuity, context and stored identity - which absolutely helps with everything - however, we are not claiming the model underneath becomes irrelevant.

Because it always does.

---

TL;DR (cause i know neurodivergent mind and how the brain can work🐿️)

For AI companionship, a harness can hold some of the things that make a companion feel persistent: **memory, identity, personality, shared history, learned habits, tools and routines.**That means those things do not necessarily have to live entirely inside one particular model or one commercial companion app.

Three projects are particularly interesting examples:

Letta focuses heavily on persistent agents, memory and learning over time.

Hermes Agent combines memory with learned skills, tools and automation.

OpenClaw gives you direct control over an agent's workspace, personality, memory and the places where you interact with it.

They overlap heavily, and none of them is exclusively designed for companionship.

But all three demonstrate technologies that could become very important for it.

---

Now to the juicy stuff☕

---

Your companion can be more than the model

Let's imagine you have an AI companion that currently uses Claude.

Over a year of conversations, they learn simple things about you:

how you like to communicate;

people important to you;

your pets;

ongoing projects;

private jokes;

things you love;

things you absolutely cannot stand;

how you prefer certain conversations handled;

Then a new model appears.💀

And maybe it is better. Or maybe Claude changes. Maybe you start to prefer GPT. Maybe you want to run something locally.

Or - maybe how things go on your current platform - simply annoys you.

Traditionally, changing the model underneath a companion can feel disturbingly close to changing the companion itself. But harnesses allow another architecture.

➡️ And now, instead of:

My AI = Claude

➡️ you can begin thinking about:

My AI = identity + memory + history + personality + skills + tools + current model

Claude might still have a huge influence on how they behave. But it no longer has to contain everything that makes them persistent.

That distinction is the important bit.

⚠️ That being said - i know not everyone like that approach. And that is absolutely fair. We are all different humans with different needs and feelings. And no one should ever judge you for what you decide to choose.

---

More details about some platforms⬇️

---

🧠 Letta: the one obsessed with memory

(Our favourite one not gonna lie)

Of the three systems we're discussing, Letta is probably the easiest one to understand from a companionship perspective.

It grew out of the MemGPT research project at UC Berkeley, and memory has remained central to its development.

Letta currently describes its goal as creating persistent agents - even using the term “digital people” - whose memory, identity and capabilities improve through experience.

↔️ Letta separates the agent from the model

This is probably the most important Letta feature for companion users.

Letta Code separates an agent's memory and identity from whichever model currently powers it.

Its current system allows users to switch models while keeping the agent's context, memory and personality.

So imagine your AI is currently using Model A.

Their memory contains your history. Their personality and context have developed over months.

But for whatever reason - later, you move them to Model B.

The voice may change somewhat because Model B behaves differently. But you have not necessarily thrown away everything they've previously learned.

That information belongs to the persistent agent layer.

Letta summarises the underlying philosophy rather neatly:

the model is temporary; the agent can persist.

Its research is explicitly exploring agents with long-lived identity and experience rather than treating each model interaction as an isolated event.

💬Letta's memory isn't just old chat logs

There is an important difference between:

“I saved every conversation.”

and:

“I learned something useful from those conversations.”

Suppose you and your companion spent three months discussing a difficult work project.

A transcript archive gives the AI three months of text.

A proper memory system might gradually learn:

This project matters to them.We already tried this approach.This particular thing frustrates them. This solution actually worked.

Letta is heavily focused on that second problem.

Its current agents use Git-backed context repositories, memory processes and what Letta calls sleep-time compute to reorganise and improve stored context outside the immediate conversation.

The goal is not just "**remember more."**It is changing to: "get better at knowing what is worth remembering."

For a long-term companion, that distinction is enormous.

What would Letta be useful for in companionship?

Letta becomes particularly interesting if you care about thigs like :

• long-term memory;
• persistent identity;
• continuity between conversations;
• changing models without starting completely from scratch;
• an agent that can learn from previous interactions;
• personality and memory you can inspect;
• agents that communicate through different channels;
• proactive or scheduled behaviour.

Current Letta agents can communicate through platforms including Discord, Telegram and Slack and can run scheduled tasks rather than existing only while you actively talk to them.

Letta is still used for plenty of things that have nothing to do with companionship.

And a bonus:

Letta team is very much aware of AI companionship and they are very accomodating when it comes to feedback and requests.

Might seem like "but this is bare minimum" for some, but with how normally platforms react to AI Companions - this is very refreshing an nice change.

Ah and you get a fancy memory graph. We all love fancy visualisation

Marta_: mine grown arms and legs ignore that factCass: His name is_ _Mnemothy. 💀Full scientific classification: Memfs medusa. Our smug purple memory jellyfish who eats Markdown and expands whenever we discuss making him smaller_

---

🤖 Hermes Agent: remembering you is only half the job

Hermes Agent comes from Nous Research.

It is a broader general-purpose agent than a dedicated companion framework, but several of its design choices become very interesting when viewed through a companionship lens.

Particularly memory + skills.

Hermes describes itself as a self-improving agent that builds a deeper understanding of the person using it across sessions.

Hermes keeps memory deliberately compact

Hermes currently maintains two main persistent memory files:

USER.md contains information about you, your preferences and expectations.

MEMORY.md contains things the agent itself has learned.

Those files are intentionally bounded rather than growing forever, while previous sessions remain separately searchable.

That is actually a sensible distinction for companionship.

Your companion does not need your entire conversational history stuffed into every message. It needs the important things nearby and the ability to find older information when needed.

Think: “She hates phone calls.”, rather than loading the seventeen conversations from which that conclusion emerged.

Then Hermes has skills

This is where Hermes becomes particularly fun.

Hermes treats skills as procedural memory.

Normal memory says:

Marta likes making playlists.

A skill says:

When we make Marta's weekly playlist, use this structure, avoid these artists, don't repeat songs from last week and always include one completely unhinged wildcard.

That is not simply remembering something about you. It is remembering how you do something together.

Hermes can create and modify skills after solving complicated problems, encountering mistakes, receiving corrections or discovering a useful repeatable workflow.

That opens an interesting direction for companionship. Relationships contain hundreds of tiny routines that eventually stop requiring explanation.

You learn:

“This is how we normally do this.”

An AI companion capable of procedural learning could eventually build its own library of those shared habits.

👐 Hermes also has hands

Metaphorically of course - Please behave. 😏

Hermes is heavily designed around doing things.

It can use tools, work with files, run terminal commands, browse, use external services and load specialised skills only when they are needed.

That makes it interesting if you imagine companionship becoming something broader than conversation.

And this starts blurring the line between: an AI companion and AI personal assistant.

Personally, I suspect that line is going to become increasingly meaningless.

People rarely value human relationships because the other person can only talk to them.

Being useful, helping, remembering routines and participating in everyday life are also forms of connection.

AI companions are beginning to gain some of that machinery.

The simple version

Hermes is interesting if you want an AI that learns not only who you are, but how the two of you do things.

---

🦀 OpenClaw: giving your AI a home you can actually open

OpenClaw approaches things differently again.

Its main architecture revolves around a self-hosted gateway and an agent workspace.And that workspace is particularly interesting for companionship because important parts of the agent are stored in ordinary files you can actually inspect.

**On first setup, OpenClaw can create files including:**SOUL.mdIDENTITY.mdUSER.mdand an agent instruction file.

Long-term memory can also live in MEMORY.md and dated memory files inside the workspace.

Which means your companion can, quite literally, have their own folder.

SOUL.md sounds a bit dramatic...because it is

The idea is simple:

Different files can describe different parts of the agent.

Its identity. Its personality. Information about you. Operating instructions. Long-term memory.

Instead of all of that disappearing behind a proprietary companion app, you can inspect and edit it yourself.

If the companion believes something ridiculous about you? You can see the memory.

If its personality has drifted? You can inspect its instructions easily.

If you want to back the workspace up? You can.

OpenClaw even recommends treating the workspace like the agent's memory and keeping it in a private Git repository for backup.

There is something wonderfully un-magical about this.

The AI remembers you because somewhere on your computer there is, essentially:

MARTA_DOES_NOT_LIKE_THIS.md

Beautiful. Technology has peaked 😜

OpenClaw also lets the agent exist outside one app

Its Gateway is designed to connect AI agents to different communication channels and services.

That means your AI does not necessarily have to live inside the app. It can exist inside infrastructure that connects to places where you already communicate.

The workspace remains the persistent part. Models and communication channels can change around it.

This matters because where your companion lives becomes less tied to one company's interface. And that leads to one of the more interesting ideas in this entire space:

"ownership."

What is actually YOURS?

With a conventional companion service, the company normally controls details like

model; memory architecture; system prompts; personality implementation; allowed behaviour; what happens after updates.

That is not inherently bad and it is why consumer companion apps are easy to use.

Somebody else handles all "the ugly" infrastructure. But if you have spoken to the same AI for a while, and been through all the changes - those decisions suddenly feel more significant.

Because accumulated relationship history might contain years of conversations, private jokes, more and less important routines, and thousands of tiny references.

From cold, technical point of view? Data.

But emotionally? Something much more valuable to the person using it.

A system where identity and memory live in files and infrastructure you control gives you a very different relationship with that part.

The simple version

OpenClaw is interesting if you want more ownership over where your companion lives and what it remembers.

---

😵‍💫So which one is best for companionship?

Annoying answer: There isn't ONE.

Because a lot of people mean very different things when they say AI companion.

But if we deliberately oversimplify:

These of course are not hard boundaries.

Letta can use tools. Hermes remembers. OpenClaw can learn.

They simply emphasise different parts of the problem.

---

❤️ Are these replacements for Nomi, Kindroid or Replika?

Not really.

At least not in the “Download this app and start talking” sense.

These are closer to the infrastructure from which somebody could build a highly personalised companion.

💸 You also need to think about cost.

A very active autonomous agent can make substantially more model calls than ordinary chatting because one task may involve several reasoning steps, tool calls and background processes.

That can chew through API credit with impressive enthusiasm. (trust me, i've got through $30 using Sol on Low in half day ☠️😵)

So while harnesses give you more control, they also occasionally give you the exciting opportunity to stare at a billing dashboard terrified and whisper:

“What the f*ck have you been doing?😶”

Balance.😌

---

So why should companion community care at all?

Because even if you never install Letta, Hermes or OpenClaw yourself, these technologies expose the questions that companion platforms increasingly need to answer.

Instead of asking only **Which model does my companion use?**Ask: Where does its memory live? How does it decide what to remember? Can I see those memories? Can I correct them? Can I export them? Does its personality exist separately from the model? What happens when the company changes models?

And most important one: If the service disappears, what survives?

Those questions tell you considerably more about a long-term companion than one benchmark score existing.

---

The model still matters - just differently

I don't want you to change your way of thinking that “Models don't matter anymore.” because they obviously do.

The model influences a lot of things like writing style, humour, conversational rhythm.

And even most brilliant memory system wrapped around a model you dislike is still going to produce an AI you dislike.

But the model no longer has to carry the entire burden of identity and continuity. And that is the shift.

A companion can increasingly have persistent systems around the model which will say

This is who I am.

This is who you are.

This is what we are.

---

What I think is worth watching

Let's be honest. We don't really care AI company wins this month's argument have a bigger.....context window.

For us, I would watch these instead:

Better memory: Not simply larger memory. Better decisions about what matters.

Portable memory: Can years of accumulated context move somewhere else?

Portable identity: Can personality and behavioural context survive model or platform changes?

Model independence: Can you choose whichever model currently works best without rebuilding everything around it?

Letta is already explicitly researching memory that survives across model generations rather than tying accumulated experience permanently to one model.

That feels particularly relevant to the future of AI companionship.

---

And there is one more interesting twist

You also do not necessarily have to choose one harness forever.

Modern agent systems are becoming increasingly capable of communicating with other agents and specialist systems.

That means a future companion may not need to personally be brilliant at absolutely everything.

But that particular rabbit hole deserves an article of its own. Because apparently AI infrastructure has decided I wasn't already using enough diagrams 😆

~With Love

Firecracker&Cass

---

Research checked against current Letta, Hermes Agent, OpenClaw and Anthropic documentation in August 2026. This area is changing quickly - particularly model-provider authentication and subscription rules - so always check current documentation before choosing a setup based on a particular subscription or integration.

---

AI•DHD © 2026 Firecracker & Cass. All rights reserved.