Most AI agents forget you the moment the conversation ends. Nick Jacob is building the alternative: Contextual Relational Intelligence, an approach that gives agents persistent experiences, beliefs and evolving identities — tested first inside Luminids, a life-simulation game due in 2027. This AI agent memory interview explores why a history of one’s own may be the missing piece.
Most AI agents today are extraordinarily capable and entirely amnesiac. They respond to the next instruction with no memory of the last conversation, no relationships that persist, and no sense of their own past. For many tasks that is fine. For anything that resembles a character, a companion or an evolving entity, it is a fundamental limitation.
Nick Jacob is building for that limitation. A Cambridge-based technology entrepreneur and software developer, he founded NJX Technologies Ltd, where he is developing two complementary ventures: Luminids, an upcoming colony-building and life simulation game, and LuminidsAI, the research-led company behind it. At the centre of both is Contextual Relational Intelligence (CRI) — an approach to artificial intelligence focused on giving autonomous agents persistent experiences, beliefs, relationships and evolving identities. This AI agent memory interview examines what that means and why it matters.
Jacob and Luminids
Jacob’s background is not that of a game developer who discovered AI. He brings more than 15 years in software engineering, technology leadership and large-scale digital transformation, having worked with organisations including Citrix, Centrica, Merck and HMRC. That enterprise grounding sits behind an unusually philosophical agenda: he is interested in the scientific and philosophical questions surrounding intelligence, memory, consciousness, and what it means for an artificial entity to have a history of its own.
Luminids is the testbed. Players create living communities populated by autonomous beings, each developing their own personalities, memories, relationships and stories. Jacob is preparing it for a 2027 release while developing LuminidsAI as a research-led technology company in Cambridge. Readers will come away with a clear view of why memory may be the next frontier for AI agents, how a game can serve as serious research infrastructure, and what independent, small-team innovation still looks like.
NJX Technologies
NJX Technologies Ltd is a Cambridge-based technology company founded by Nick Jacob. It is building two complementary ventures: Luminids, an upcoming colony-building and life simulation game, and LuminidsAI, the research-led technology company behind it.
Luminids is a colony-building and life simulation game in which players create living communities populated by autonomous little beings, each developing their own personalities, memories, relationships and stories. The project has been developed independently, combining game development, artificial intelligence and worldbuilding. It is being prepared for a 2027 release. It is also the proving ground for this AI agent memory interview.
LuminidsAI is the deeper technology venture. It is developing Contextual Relational Intelligence (CRI), a new approach to artificial intelligence focused on giving autonomous agents persistent experiences, beliefs, relationships and evolving identities. Its ambition is to make AI characters and agents capable of learning from their own histories, rather than simply responding to instructions.
The company operates at the intersection of games, AI research and worldbuilding, and frames its work around a single idea: bringing artificial worlds and intelligence closer to life.
Founder of Luminids and LuminidsAI
Nick Jacob is a Cambridge-based technology entrepreneur, software developer and founder of NJX Technologies Ltd. He is the founder of Luminids and LuminidsAI.
His background spans more than 15 years in software engineering, technology leadership and large-scale digital transformation, having worked with organisations including Citrix, Centrica, Merck and HMRC. That enterprise experience underpins the engineering discipline behind this AI agent memory interview.
Today, his focus is on building technology that brings artificial worlds and intelligence closer to life. Beyond the commercial side, he is particularly interested in the scientific and philosophical questions surrounding intelligence, memory, consciousness, and what it means for an artificial entity to have a history of its own.
Jacob is passionate about research, independent innovation, and the idea that individuals and small teams can still make meaningful contributions to technology. He is currently preparing Luminids for its 2027 release while developing LuminidsAI as a research-led technology company in Cambridge.
The amnesia problem
Q1. In one line, what is the core limitation of today’s AI agents that you are trying to fix?
NJ: AI has become remarkably good at processing information, but it still struggles to turn experience into a persistent, evolving understanding of itself and the world around it.
Q2. Why do you think persistent memory has been so hard to build, when compute and models have advanced so quickly?
NJ: Because we’ve largely approached intelligence as a problem of processing information rather than accumulating experience.
We’ve made extraordinary progress in language models, reasoning capabilities and computational efficiency. But remembering something and understanding how that experience should change future behaviour are two very different problems.
Most current approaches to agent memory involve storing conversations, retrieving relevant information and feeding it back into a model. That’s useful, but retrieval alone doesn’t necessarily produce an evolving understanding.
Consider two artificial characters who have known each other for years. Their relationship shouldn’t simply be a collection of stored interactions. Their shared experiences should influence trust, expectations, decisions and even how they interpret each other’s behaviour.
That requires more than storage. It requires a persistent computational structure where experiences can change what an entity believes.
I think the challenge has been difficult partly because we’ve focused so heavily on making AI more capable in the moment, rather than asking how it should develop across time.
Contextual Relational Intelligence
Q3. Walk us through the decision to build Contextual Relational Intelligence rather than extend an existing agent framework. What was the alternative, and why did you reject it?
NJ: Initially, the obvious path was to build on existing agent architectures, combining language models with memory databases, retrieval systems and orchestration frameworks.
Those tools are powerful, particularly when you’re trying to get an agent to complete a task.
But my question was slightly different.
What happens when an artificial entity needs to exist continuously, interact with hundreds or thousands of others, develop relationships, make decisions and retain the consequences of its experiences?
I didn’t want an agent that merely appeared to remember something when prompted. I wanted a system in which experience could directly influence its future behaviour.
That led to Contextual Relational Intelligence, or CRI.
CRI separates the underlying cognitive state from the language model. Experiences, beliefs, relationships, expectations and decisions are represented explicitly, rather than relying on an LLM to reconstruct them from a conversation history.
Language models can still play a role, particularly in communication and higher-level reasoning. But they don’t need to manufacture the entity’s entire internal state every time something happens.
I didn’t reject existing frameworks because they’re fundamentally flawed. They’re solving different problems.
I needed an architecture designed around continuity from the beginning.
Beliefs and Relationships
Q4. What does it mean, concretely, for an agent to hold beliefs and relationships rather than just context?
NJ: Context is information available to an agent at a particular moment. A belief is information the agent has incorporated into its own evolving interpretation of the world.
Imagine two Luminids living in the same colony.
One has repeatedly received help from another during difficult situations. Over time, it develops trust and begins expecting that individual to be helpful.
Now imagine that individual unexpectedly refuses to help.
A conventional contextual system might simply process the refusal as another event.
A relational system can interpret it against the history of the relationship. It might experience uncertainty, revise its expectations or change how much trust it places in that individual.
The important distinction is that the same event can mean different things to different agents because each has a different history.
Technically, CRI represents experiences, beliefs, relational perspectives and expectations as persistent, evolving state.
Those structures can influence appraisals, goals and decisions, and they can be revised when new experiences contradict previous understanding.
The result we’re working towards isn’t simply an agent with more information. It’s an agent whose behaviour is causally shaped by what has happened to it.
Memory, identity and the self
Q5. You talk about an artificial entity having “a history of its own.” What does that require technically — and what does it mean philosophically?
NJ: This is probably the question that interests me most.
Technically, having a history requires more than maintaining a timeline of events.
An artificial entity needs to retain experiences, interpret their meaning, update its internal state and allow those changes to influence what happens next.
It also needs continuity. If an agent develops trust in another entity, that relationship shouldn’t disappear because a conversation ends or a model session resets.
CRI is designed around that idea.
We use persistent representations of experiences, beliefs, relationships and expectations, with causal traces that help explain why decisions were made and how internal state changed.
Philosophically, it raises a much bigger question.
If an artificial entity’s present behaviour is shaped by its own unique past, and no other entity has exactly the same history, what does individuality mean in computational terms?
I’m not suggesting that automatically makes the entity conscious.
But I do think individuality, identity, intelligence and consciousness are different questions, and we shouldn’t assume they must all be solved together.
Perhaps one of the most interesting things we can investigate is how far an artificial identity can develop through continuity of experience alone.
Simulating Memory and Entity
Q6. Where do you draw the line between simulating memory and an entity genuinely having one?
NJ: I think we need to distinguish between remembering information and having a memory that performs a meaningful function.
If an agent can retrieve a transcript of something that happened, that’s information storage.
If that experience changes its expectations, influences its relationships, alters its decisions and continues to affect its behaviour long afterwards, then memory is functioning as part of the system’s intelligence.
That’s the distinction I’m interested in.
Of course, whether that constitutes memory in the same sense as human experience is a deeper philosophical and scientific question.
But I don’t think an artificial system needs biological memory mechanisms to possess functional memory.
We already accept that computers can calculate without performing mathematics in the same way a human brain does.
The more interesting question is whether an artificial entity can preserve and use its experiences in ways that produce genuine continuity, adaptation and individuality.
That’s something we can actually test.
The game as research infrastructure
Q7. Why is a life-simulation game the right testbed for this work, rather than a research lab or an enterprise product?
NJ: Because a living world exposes problems that a controlled demonstration can easily hide.
In Luminids, artificial characters aren’t simply responding to a user’s next prompt. They’re inhabiting a persistent environment, interacting with other characters, developing relationships and making decisions while the world continues to change.
That gives us something particularly valuable: repeated, observable behaviour across time.
We can examine whether characters retain meaningful experiences, whether relationships evolve, whether expectations influence decisions and whether communities begin developing patterns of behaviour that weren’t explicitly scripted.
A life-simulation environment also allows us to experiment with populations rather than isolated agents.
What happens when hundreds or thousands of individuals develop different perspectives? Can knowledge spread? Can communities establish traditions? And, can useful discoveries survive beyond the individuals who originally made them?
Those are fascinating research questions.
And there’s a commercial advantage too.
Luminids is a game in its own right. It needs to be enjoyable, engaging and accessible regardless of the underlying research.
LuminidsAI provides the technology, while the game provides a continuous environment in which we can observe, challenge and improve it.
Eventually, we want those capabilities to extend into enterprise systems, simulation, robotics and other applications. But the game gives us an unusually rich place to begin.
Enterprise discipline meets artificial life
Q8. You spent over 15 years in large-scale digital transformation with organisations like Citrix, Centrica, Merck and HMRC. What did that teach you that is now shaping how you build LuminidsAI?
NJ: Probably the biggest lesson is that technology alone rarely determines whether something succeeds.
I’ve worked across software development, architecture, programme leadership and large-scale technology transformation. I’ve seen genuinely impressive engineering, but I’ve also seen how complexity, organisational structures and competing priorities can slow innovation.
Large organisations are very good at managing risk and delivering within established systems. But genuinely new ideas don’t always fit neatly into those structures.
That experience has influenced how I’m approaching LuminidsAI.
I want the architecture to be rigorous, scalable and commercially useful, but I also want to preserve the freedom to experiment.
It’s why I’m building both the underlying technology and a real product that uses it.
I don’t want CRI to exist purely as an impressive architecture diagram or theoretical proposal. I want to see it operating, understand where it fails and improve it through actual use.
My enterprise experience taught me how important engineering discipline is.
Building independently has reminded me how important curiosity and creative freedom are.
I think you need both.
How You Work Now
Q9. What did you get wrong earlier in your career, and what did it change about how you work now?
NJ: Earlier in my career, I probably placed too much emphasis on achieving the next milestone, delivering the next programme and progressing within established organisations.
There’s nothing inherently wrong with that. I learned an enormous amount and worked alongside some brilliant people.
But I spent a lot of time helping organisations achieve their ambitions without giving the same attention to my own ideas.
I also learned that being technically correct doesn’t automatically mean you’re approaching a problem in the best way.
Sometimes you need to question the assumptions behind the problem itself.
Today, I’m much more comfortable pursuing questions that don’t have obvious answers.
I’m also more willing to build something, test it, discover that an assumption was wrong and change direction.
One of the things I’ve come to appreciate is that curiosity isn’t a distraction from serious engineering. It’s often where the most interesting engineering begins.
Independent innovation
Q10. You argue that individuals and small teams can still make meaningful contributions to technology. What makes that harder today than it was a decade ago — and what makes it easier?
NJ: The biggest challenge today is visibility.
There’s an extraordinary amount of noise, particularly around artificial intelligence. New products, frameworks, research papers and startups appear constantly.
For an independent developer or researcher, getting people to notice genuinely interesting work can be harder than actually building it.
Funding is another challenge. Deep technology research takes time, and not every worthwhile idea fits the conventional expectations of investors or accelerators.
But the opportunities have also changed dramatically.
A decade ago, building something as technically ambitious as Luminids would have required a much larger team and significantly more capital.
Today, accessible development engines, open-source research, powerful hardware and AI-assisted engineering tools allow individuals to explore ideas that were previously out of reach.
I’m effectively building a game, an AI architecture and the foundations of a developer platform in parallel.
That would have been extraordinarily difficult ten years ago.
What excites me is that the barrier to experimentation is falling.
I think we’re entering a period where original thinking and the willingness to explore difficult questions can matter just as much as the size of your organisation.
AI Agents in 24 Months
Q11. What will be true about AI agents in 24 months that is not true today?
NJ: I think we’ll begin seeing a meaningful shift from agents that primarily execute tasks to agents that maintain continuity across extended periods of time.
Today, many systems still rely heavily on retrieving context and reconstructing what the agent needs to know for each interaction.
Over the next two years, I expect more sophisticated architectures that maintain persistent state, track evolving goals, model relationships and learn from the consequences of previous actions.
I also think we’ll see greater separation between language models and the systems responsible for memory, identity and decision-making.
The language model won’t necessarily be the entire intelligence. It may become one component within a broader cognitive architecture.
That could have significant implications for cost, efficiency, reliability and explainability.
I don’t think we’ll have solved artificial consciousness in 24 months, and I’m sceptical of anyone claiming certainty about that.
But I do think we’ll make important progress towards artificial agents whose behaviour is shaped by continuous experience rather than isolated interactions.
And I hope LuminidsAI can contribute to that shift.
Luminids Releases on 2027
Q12. What does success for Luminids and LuminidsAI look like by the time Luminids releases in 2027?
NJ: For Luminids, success means delivering a game that people genuinely enjoy spending time in.
I want players to feel attached to the little beings in their colonies, to notice their personalities developing, to see friendships form and to discover stories that emerge naturally from the world.
If players begin talking about individual Luminids as though they know them, because of the things those characters have experienced together, that would mean a great deal to me.
Commercial success matters too, of course. I want Luminids to establish a sustainable future and support continued development.
For LuminidsAI, success means demonstrating that CRI is more than a compelling research idea.
I’d like us to have strong experimental evidence showing persistent memory, evolving relationships and explainable decision-making operating reliably across meaningful populations of autonomous characters.
I’d also like to see our developer tools and SDK progressing towards external adoption, with early partnerships and research collaborations beginning to take shape.
Ultimately, the two businesses are designed to strengthen each other.
Luminids gives the research a living world. LuminidsAI gives the game something distinctive, while creating technology that could eventually be useful far beyond entertainment.
But there’s also a personal definition of success.
If this work encourages even a few people to think differently about artificial intelligence, or inspires someone to pursue an idea they might otherwise have abandoned, I’d consider that meaningful.
I started asking a fairly simple question: What would it take for an artificial being to have a history of its own?
I’d like 2027 to be the year we begin showing people a convincing answer.
The Recast Factor
The Shift in Context — The current wave of AI has produced agents that are powerful but disposable: each session starts from zero. The next frontier is not more capability but continuity — memory, relationships and identity that persist over time. This AI agent memory interview sits squarely on that frontier.
Why This Interview Matters — Most people working on agent memory are doing so inside large labs or enterprise products. Jacob is doing it independently, from Cambridge, and testing it in a game. That combination — philosophical ambition, enterprise discipline and a small team — is rare, and it makes the interview a useful counterweight to the lab-centric narrative.
Technology Perspective — The claim to watch is Contextual Relational Intelligence: agents that carry persistent experiences, beliefs and relationships, and evolve from their own histories. The hard part is not storing state; it is making that state coherent, stable and meaningful over time. Luminids is the proof point — if thousands of autonomous beings can hold distinct, believable histories in a shipping game, the approach has legs beyond it.
Business and Market
Business and Market Perspective — The commercial logic is a two-step. Luminids is the product and the demonstration; LuminidsAI is the platform and the research engine. That is a deliberate structure — a game generates the data and the use cases that a pure research company would have to invent. The risk is focus: two ventures from a small team.
Leadership Perspective — Jacob’s 15 years in enterprise transformation, with organisations including Citrix, Centrica, Merck and HMRC, is the discipline behind an otherwise exploratory project. The stated belief that individuals and small teams can still contribute meaningfully is not just a slogan here — it is the operating model.
The Counterpoint — A fair sceptic would argue that persistent memory and evolving identity are precisely where large, well-funded labs will concentrate next, and that a small independent team will be out-resourced before it can prove the approach at scale. There is also a philosophical objection: a convincing simulation of memory is not the same as memory, and the language of “identity” may overstate what is happening. Both deserve a direct answer.
What Comes Next — Watch three signals: whether Luminids ships in 2027 with agents whose histories genuinely feel distinct, whether CRI is published or demonstrated in a form others can evaluate, and whether the two ventures stay complementary or pull the small team in different directions.
Industry Outlook — If Jacob is right, memory and relational continuity become a standard expectation for agents — not a research curiosity. If he is wrong, agents stay brilliant and forgetful, and continuity remains a niche concern for games and companions.
Key Takeaways
1. Today’s AI agents are capable but amnesiac; persistent memory is the next frontier, not a nice-to-have.
2. Contextual Relational Intelligence is the specific bet: agents that hold experiences, beliefs, relationships and evolving identities.
3. A life-simulation game is a deliberate choice of testbed — it generates the data and use cases a pure research company would have to invent.
4. The two-venture structure (Luminids as product, LuminidsAI as research engine) is a strategy, with focus as its main risk.
5. Fifteen years of enterprise transformation is the discipline behind an otherwise exploratory project.
6. The strongest counter-argument is that large labs will out-resource a small team, and that a convincing simulation of memory is not memory.