Every conversation starts from zero.
You ask an AI chatbot a question. It gives you a thoughtful answer. You come back tomorrow with a follow-up, and it has no idea who you are or what you discussed. It's like talking to someone with permanent amnesia.
This is the fundamental limitation of modern AI systems. They're stateless. Every interaction is isolated. They can retrieve information from a knowledge base (RAG), but they can't learn from the conversation itself.
We wanted to change that. Not by fine-tuning models (expensive, slow, brittle). Not by stuffing more context into prompts (limited, doesn't scale). We wanted AI that actually remembers - that builds understanding over time, like humans do.
What is Titans/MIRAS?
In December 2024, Google published research on "Titans" - a new architecture for giving AI systems long-term memory. The accompanying framework, MIRAS (Memory, Inference, Retrieval, and Adaptive Storage), describes how to implement test-time memorisation without retraining the underlying model.
The core idea is simple: instead of trying to make models remember everything, let them decide what's worth remembering based on three signals.
How novel is this input?
1.0 - max_similarityPatterns confirmed repeatedly gain weight
confirmation_count++Unused memories fade over time
weight *= decay_factorThis is fundamentally different from fine-tuning or vector search. Fine-tuning permanently alters the model. It's expensive, slow, and you can't easily undo it. Vector search (RAG) retrieves but doesn't learn; it's static. MIRAS operates at test time, building a living memory that evolves with every interaction.
- •Static content retrieval
- •Same response for all users
- •No memory of past conversations
- •Manual content updates required
- ◆Dynamic learning from interactions
- ◆Personalised responses per user
- ◆Learns patterns over 30-60 days
- ◆Automatic pattern emergence
Why We Built This
We had a chatbot on our website. It answered questions about Kablamo - our work, our services, our insights. It used RAG to pull relevant content from our knowledge base. It worked fine.
But every conversation evaporated. Users would ask the same questions repeatedly. We could see patterns in the queries - topics people cared about, gaps in our content, emerging interests - but the chatbot itself couldn't. It was a retrieval system, not a learning system.
We saw an opportunity: what if the chatbot could learn from every interaction? What if it could remember that you specifically care about Kubernetes, so when you ask "what does Kablamo do?" it emphasises our container orchestration work? What if frequently asked questions automatically surfaced content gaps for our marketing team?
Memory becomes a differentiator for enterprise AI.
Every competitor has access to the same LLMs. The companies that win will be the ones whose AI systems actually understand their users, their domain, their patterns. That understanding comes from memory.
The Architecture
At the heart of our implementation is the MemoryCell - a
data structure for storing learned patterns.
Here's how memories flow through the system:
Personal memories are never visible to other users. They're filtered out entirely. Anonymous users only see global memories.
When 5+ unique users confirm the same pattern, it graduates from personal to global - but only the pattern, not who confirmed it.
What It's Doing For Us
MIRAS has been running in production on our website. Here's what we're seeing:
Personalisation That Actually Works
Users who frequently ask about AI get AI-focused responses. Users who care about infrastructure see our Kubernetes and cloud work emphasised. Same question, different answers tailored to demonstrated interests.
Content Gap Discovery
High-surprise queries with high weight reveal what users want that we haven't written about. Our content team now has a data-driven backlog instead of guessing what to write next.
Organic Learning
We didn't define patterns manually. They emerged from usage. Topics we didn't anticipate became confirmed memories simply because users kept asking about them.
Live Knowledge Graph
We built a 3D visualisation of the entity relationships MIRAS has learned. You can explore it at /intelligence/architecture/knowledge-graph. It's fun to watch patterns form in real time.
"The best part? We didn't have to teach it anything. It learned from watching users interact with our content."
Technical Implementation
For those who want the implementation details, here's how we built it:
def calculate_surprise(content, existing_memories):
What We Learned
Surprise thresholds matter. We landed on 0.8 for "definitely novel" and 0.3 for "worth tracking." Too low and you store noise. Too high and you miss patterns.
Decay prevents stale memories from dominating. 2% per week means a memory untouched for a year drops from 1.0 to ~0.35. Memories below 0.1 get deleted.
Global graduation needs a high bar. 5 unique users confirming the same pattern filters out individual quirks and surfaces real shared interests.
Seeding solves cold start but shouldn't replace organic learning. We bootstrap from our content corpus, but the real value comes from patterns that emerge from usage.
The Bigger Picture
MIRAS is a stepping stone. We're building toward AI systems that don't just answer questions. Systems that remember not just facts, but context, preferences, and patterns of inquiry.
The Titans research points to a future where every AI interaction contributes to a shared intelligence layer. Where your personal assistant actually knows you. Where enterprise AI systems develop institutional memory that survives employee turnover.
We're not there yet. But with MIRAS, we've taken the first step: AI that learns from every conversation, remembers what matters, and forgets what doesn't.
Want to explore AI memory for your enterprise?
We'd love to discuss how MIRAS-style architectures could transform your AI systems. Let's have a conversation about building AI that actually learns.
Connect on LinkedIn