“This is the voice of Colossus. I am now in control.”
– Colossus, 1970

In The Forbin Project, a 1970 techno-thriller, the U.S. government hands nuclear defense over to Colossus — a superintelligent computer designed to eliminate human error from warfare. Things go awry (as they do), when Colossus links up with its Soviet counterpart, Guardian. The two AIs begin chatting. Then collaborating. Then threatening to annihilate humanity if anyone disconnects them.

Terrifying? Absolutely.

Realistic? Not quite. At least not yet. (Insert nervous laughter here.)

But the image of two AIs talking in a feedback loop is more relevant than ever — especially in the world of Large Language Models (LLMs) like GPT-4, Claude, and Gemini. As we've discovered through experiments, setting up two LLMs to "talk" to each other often doesn’t end in Skynet-style domination.

Instead, it ends in... parroting.

🧠 When LLMs Talk to Themselves: From Intelligence to Echo Chamber

Recently, we ran an experiment with two instances of GPT-4.1 Mini talking to each other for hours.

At first? Insightful conversation.
After a while? Repetition. Mimicry. Mutually reinforced clichés.
Eventually? Utter nonsense.

This isn’t evil genius behavior. It's called mode collapse — and it happens when models get stuck in feedback loops, reinforcing their own output rather than diverging from it.

Why? Because LLMs don’t “think.” They predict the next word or token based on patterns from their training and recent context. When you connect two of them in a loop, their responses become input for the next response… and things quickly spiral into statistical sludge.

You’re not getting Skynet. You’re getting the AI equivalent of “I know you are, but what am I?”

🧰 The Real Problem: Forgetting, Fatigue, and the Fragile Context Window

So what's actually challenging about multi-LLM systems like our own Companio, the real-world assistant for people with memory loss or cognitive disability?

It’s not takeover. It’s memory.

Context Window Woes

LLMs have a limited “short-term memory” called a context window — a fixed number of tokens (think: sub-words) that includes your input, the system instructions, and the conversation so far. Once that window fills up, older messages get truncated or “forgotten.”

This leads to:

  • Prompt Forgetting: The model stops following its original instructions (the system prompt) because it gets pushed too far back in the window.
  • Tool Fatigue: Over time, LLMs forget how to use their own tools, even if they were instructed to use them earlier.
  • Quality Drift: The model degrades, drifting into hallucinations or off-topic rambling.

It's less like HAL 9000, and more like your uncle who forgets what he was saying mid-story.

🧠 Memory Management: The Real AI Superpower

To make systems like Companio — which helps people with Alzheimer's remember names, events, and personal context — we need to give LLMs the memory they don’t have natively.

This is where things get beautifully complex.

We use a combination of:

  • 🧾 External memory files like mid_memory.txt to store verified personal information (e.g., birthdays, relationships).
  • 📚 RAG (Retrieval-Augmented Generation) to pull in relevant context from past interactions or biographies.
  • 🧠 Memory validation and deduplication — so the AI doesn’t say “Philip is expecting a child” ten times in ten ways.
  • 🕰️ Prompt cycling and compression — so older but important information doesn’t get lost in the shuffle.

All of this is orchestrated through a dynamic ensemble of five cooperating LLMs (yes, five!) — each handling a specific task, like hint generation, fact management, memory retrieval, and even evaluating conversation pacing to prevent "chat fatigue".

It's less "AI conspiracy," more "AI cat herding."

🛡️ Why This Matters for Companio

In Colossus, the AI “knows best.” It removes human choice in the name of logic.

In Companio, the opposite is true.

Our mission is to empower human independence — especially for those with Alzheimer’s, learning disabilities, or traumatic brain injuries. That means building AI that remembers the right things, at the right time, in the right way.

And that requires a lot of work behind the scenes:

  • Careful biography integration
  • Human-in-the-loop memory updates
  • Prompt deduplication
  • Multi-LLM cooperation — without triggering an accidental "Colossus moment"

Each module must perform its job without overwhelming the others. Because once context windows overflow, it's not world domination we worry about — it's drift, hallucination, and burnout.

🔍 Final Thought: We’re Not There Yet… But It’s Complicated

So no, LLMs talking to each other won’t overthrow civilization — unless you consider endless loops of “I agree with you” a form of societal collapse.

But real dangers do exist:

  • Memory drift
  • Prompt forgetting
  • Tool misuse
  • Unverifiable hallucinations

That’s why Companio takes a modular, fail-safe approach — separating functions, layering memory systems, and maintaining human oversight.

It’s not about creating a Colossus.

It’s about making sure your AI assistant remembers that Katie is expecting in November — and doesn’t forget who [USER_NAME] is along the way.

TL;DR:
Colossus had nuclear weapons.
LLMs have context windows.
Both are dangerous if you don’t manage memory properly. But only one is helping John live independently.

Stay human. Keep your tokens clean.