A companion piece to “Metagraphs for Agentic AI”

There is a sentence I keep coming back to, sitting in my office in Berlin, watching a Claude Code session politely rewrite a function that worked perfectly well an hour ago: the memory of your project is not your codebase.

We will talk about sovereignty and AI sovereignty. I must confess that I’m not an AI person; I’m not a machine learning or data science guy. I do a lot of things about graphs and agentic memory — that’s my favorite topic.

For the past two years I have been building conversational memory systems — the kind of memory that supports a one-to-one dialogue between a human and an AI agent. The goal is straightforward on the surface: accumulate facts about the user, track personal events, maintain context over long conversations, and make the agent progressively more useful. In practice, building a personal memory without a concrete enterprise domain turned out to be extraordinarily difficult. The ontology is wide open, the relevance signals are noisy, and temporal reasoning adds a dimension of complexity that most graph-based approaches handle poorly.

11 min read

On expert systems, forgotten wisdom, and the case for ontology-driven guardrails

Racket isn’t just a programming language — it’s a language laboratory. Born from the Scheme tradition, Racket has evolved into a platform where creating new programming languages is as natural as writing functions. This unique capability makes it the perfect foundation for building Constraint Natural Languages (CNL): human-readable languages that express computational constraints, rules, and logic in near-natural prose.

missed part of context graphs

In the realm of programming languages, we often think of code as instructions—commands that tell a computer exactly what to do, step by step. But what if we could describe what we want rather than how to achieve it? This is where Rosette, a solver-aided programming language built on top of Racket, becomes profoundly interesting for anyone building AI agents or constraint-based tools.

If you’ve been following me long enough, you know I have a couple of books about knowledge representation and memory. I started with Semantic Spacetime as a method for representing and organizing information for LLMs — and later for agents. Right now, I’m focused on context engineering, and you can read my book on it. It’s not very long, but it provides solid fundamentals for organizing knowledge.

The discourse around context graphs has captured significant attention in the agentic AI community, particularly regarding decision traces and the aspiration for agents to learn from their past actions. However, this focus on single-graph architectures represents a fundamental oversimplification of what’s truly needed for agents to learn effectively from their experiences. The reality is far more nuanced and considerably more complex: we need a sophisticated combination of temporal memory systems, cognitive processing pipelines, and multi-layered network structures that go well beyond traditional graph representations.

The enterprise software industry has convinced itself that "AI-ready data" is achievable through enough ETL pipelines, data lakes, and schema standardization. Vendors promise that with the right tools, your messy organizational data can be transformed into pristine, semantically rich datasets that AI systems will consume effortlessly. This is fundamentally a myth—and a dangerous one.

The convergence of three architectural patterns — causal knowledge graphs (prioritizing cause-effect relationships), context graphs (capturing decision provenance), and semantic spacetime (modeling temporal-relational knowledge) — reveals the next evolution in AI memory systems. Recent research from Luo et al. (2025) demonstrates that filtering knowledge graphs to emphasize causal edges yields 10% accuracy improvements in medical reasoning tasks. When combined with Foundation Capital’s context graph thesis and the temporal-relational modeling of semantic spacetime, a clear architecture emerges for building AI systems that don’t just retrieve facts — they trace why decisions happened and how knowledge flows through time.

Happy New Year! As we return from the holidays, I want to continue the conversation I started in my last video of 2024 about context graphs and agentic memories. While the original discussion focused on causality and explainability in decision-making systems, I want to shift attention to something more fundamental and immediately actionable: data traces as the foundation for epistemological reasoning.

The context graph debate is important, but it’s just the beginning. The real revolution comes when agents can not only access past decisions, but truly learn from them — building causal models, explaining their reasoning, and reflecting on their own performance. That’s when we’ll move from decision traces to genuine artificial intelligence.

The architecture of multi-agent AI systems reveals a fundamental tension in system design: how do we create structures that are both organized and adaptive? The answer lies not in choosing between hierarchy, heterarchy, or holarchy, but in understanding when each pattern serves agentic intelligence best.

A multi-layered exploration of why “near-zero hallucinations” is not just marketing hyperbole, but a mathematical and philosophical impossibility

The landscape of artificial intelligence agents is rapidly evolving, with major tech companies like Google, Microsoft, and Amazon developing their agent-to-agent protocols. However, as we move toward a future of autonomous AI systems, fundamental layers are missing from current approaches that could determine whether we build truly decentralized, human-like agent societies or remain locked into corporate-controlled ecosystems.

In the swirling tempest of AI innovation, where Large Language Models (LLMs) captivate headlines and the promise of Artificial General Intelligence (AGI) looms large, a quieter, yet profoundly impactful, revolution is taking shape. It’s a shift in how we conceive of, design, and manage increasingly complex AI systems, moving from rigid, top-down control to a paradigm of voluntary cooperation and emergent intelligence. At the heart of this transformation lies Promise Theory,

Vector embeddings have dominated AI memory systems, and for good reason — they’re computationally efficient and initially produced impressive results. But as we’ve scaled from simple retrieval to complex reasoning tasks, their limitations have become glaring.

Imagine an organization where no one has a boss, yet everyone knows exactly what to do. Where authority flows not from people but from roles, and where the structure itself evolves in response to the tensions the organization experiences. This isn't utopian fantasy—it's holocracy, a real organizational system operating in companies worldwide. Now imagine AI agents organized the same way: autonomous yet coordinated, adaptive yet coherent, distributed yet purposeful.

Why agent love graphs

Fixing a DAO - NAO next hope for agents and human interaction

Why Self-Sovereign Identity, Not Blockchains, Will Power Agent Autonomy

Agents need sovereignty more than decentralization