Essays on user-owned agent memory, personal infrastructure, and building systems that restore sovereignty in an age of AI, crypto, and complexity.
My old agentic stack was one agent per session with me in the loop for every task. Ateles is the standing fleet that replaced it. A personal agent swarm defined in Neotoma itself, where agents, tasks, and plans are all entities, and every action is attributed, auditable, and reversible.
Three viral posts agree that durable state is what makes agent loops work, then hand the job to a markdown file. What loop state actually requires once you run more than one loop.
Agent-led evaluation tells a user whether Neotoma fits. The new issues subsystem lets the same agent file a bug or enhancement, exchange messages with the maintainer, and watch the change arrive, all through MCP, without a human ever opening a browser tab.
The first problem with running multiple agents is that they forget. The second is that they can't coordinate. A state layer that solves storage first and then adds signaling becomes a nervous system, not just memory.
Async specialists can each be right inside their own domain and still ship something incoherent across domains. Some role has to reconcile cross-disciplinary conflict against explicit commitments rather than hallway consensus, and that work needs durable write integrity instead of team lore or vibes.
Software teams used to spend most of their time in execution—the middle between foundation and review. As that middle gets absorbed by capable AI, human leverage shifts to the ends: sharper standards upstream and denser judgment downstream. That is structurally positive—but emotionally hard for people whose craft lived in the middle.
Startup hiring advice fights over speed: hire slow versus hire ahead of the curve. Both assumed hiring tracks execution backlog. When execution is AI-assisted, the binding constraint is different: one human's attention on what goes into models and what comes out of them.
Humans at the ends, AI in the middle is a clean diagram. Real teams are not clean diagrams. Generalist surfaces, hybrid roles, hardware timelines, and catastrophic-blast partitions each shift where weight sits, without changing whether the ends carry it.
Specialists at the ends produce deep artifacts in their own vocabularies; the old org used the execution middle to translate in real time. AI can sit between domains as an async translation layer—replacing most operational meetings, if the underlying artifacts stay trustworthy.
When many agents write to one memory, you need proof on every row: who signed, what they were allowed to do, and no shared bearer secrets. Neotoma integrates AAuth so each agent signs with its own key and writes through grant-scoped admission.
On the PM role, AI, and the strange decision to define your job around the skill that just got commoditized.
Fourteen years ago I wrote a postmortem for Plancast. Reading it again now, the failure modes weren't really about product mechanics; they were about missing personal AI agents, a proper substrate for them to write to, and a sovereign mesh between them. All three exist now, and the original mission becomes tractable in a way the feed era never could.
Most teams bolt agent memory onto whatever database they already have. It works until two agents write to the same store. Then one bad write propagates at machine speed, triggering downstream actions before any human can intervene. The industry is heading toward a trust crisis that retrieval optimization won't fix.
Three independent AI agent platforms worth billions converged on plain text files for memory. The convergence validates the problem. The failure modes they share define what comes next.
OpenClaw stores memory in markdown files. Neotoma v0.4.3 plugs in natively as a structured state layer underneath, giving OpenClaw agents provenance, entity resolution, and versioned history without replacing the agent.
Sarah Wooders argues memory is the harness. She's right about context management. She's wrong that context is the whole problem.
The metrics that drive adoption in AI memory are almost all retrieval metrics. Good retrieval is necessary. No widely used benchmark tests what happens to stored data after agents write to it.
BEAM tests retrieval at 10 million tokens. State integrity degrades at 500K. The two failure modes activate at different scales, and nobody benchmarks the earlier one.
About eighteen of twenty-six people I reached ran the same evaluation prompt through their AI tools. The writeups were sharper than any call. That feedback reshaped the product’s positioning — and then its acquisition flow. The homepage now asks agents to evaluate, not humans to sign up.
I shipped eleven Neotoma releases in the first five weeks after the developer release. The CLI now works on other machines, unblocking onboarding. The MCP server is stable enough for daily agent use. The database can survive operator mistakes. Here is what changed and why.
Eighteen human product evaluators ran the same evaluation prompt through their AI tools. The writeups were sharper than any call. That feedback reshaped the product’s positioning — and then its acquisition flow. The homepage now asks agents to evaluate, not humans to sign up.
The debate over AI maps onto a question the Enlightenment never resolved. Rousseau said tools corrupt us. Condorcet said they perfect us. Nietzsche said both miss the point. Technology is a test of character, and the only honest answer is what you become in the encounter.
I accidentally overwrote my production Neotoma database and dropped from 6,174 observations to 84. I got nearly all of it back. The recovery worked because Neotoma's architecture stores observations as an append-only log and recomputes entity state from that log.
Narrative FOMO followed me from crypto to AI. At Leather, the Ordinals rush felt like real demand. At Neotoma, every memory-related tweet triggers the same scramble. I've lived both ways a narrative can break your product. Borrowed signal has a term limit.
Claude and ChatGPT now offer memory for free, but what they store are profile snippets, not the details of what you actually worked on. My testing showed stale exports, one-shot access, and platform-gated portability. That is why I am building a truth layer underneath.
I joined Blockstack in 2018 for the off-chain developer tools. Over seven years I watched token economics reshape the product direction, push the perceived moment of value perpetually into the future, and replace empirical product development with narrative-driven roadmaps.
Creation, whether a startup or a blog post, is a continuous habit of revealing your inner world and tuning into what comes back. The MVP is a continuum. So is everything else.
The structural pressures that underpin my work, and what would invalidate them as the AI industry evolves.
I overhauled the Neotoma site. The old single-page wall of text is now a visual presentation backed by documentation, tool-specific integration guides, and architecture deep dives—driven by what testers said and where they got stuck during the developer release.
The people who embrace AI encounter its problems more directly than anyone. But they also have the most faith that humans can work through them. The real skepticism isn't about the machines.
I built an agentic stack to dogfood Neotoma and accelerate my work. A private monorepo with 12+ MCP servers where I cowork with AI agents daily. Neotoma provides the structured memory underneath, letting agents build on prior work across every session.
A new open-source memory agent from Google ditches vector databases for LLM consolidation, trading determinism for active pattern discovery. It makes opposite architectural choices from a truth layer on every axis. The strongest architecture may combine both: consolidation above, deterministic state you can audit and revert below.
Structured agent memory you can inspect, replay, and trust. Open-source, local-first, available via npm for ChatGPT, Claude, and Cursor.
Retrieval fits coding and is cheap; it breaks on completeness, consistency, and provenance. Why convergence to structured state is hard, and how Neotoma is designed for it.
Command centers for agents need a single, durable state layer for tasks and visibility. The UI is the dashboard; the layer it reads and writes is the substrate.
The humanness isn't in typing every word. It's in deciding what's worth saying.
The future of crypto wallets is agentic. Wallets that expose execution surfaces let agents monitor, reason, and execute within policy while you keep sovereignty and approve only what exceeds your limits.
Retrieval for agent memory should be driven by structure, not similarity. Learned hierarchies beat RAG but are brittle. Schema-first design gives the same advantage without putting the LLM in the critical path.
You don't protect it through secrecy or patents. You build something they structurally can't pursue.
OpenClaw runs on your machine, clears your inbox, and builds skills. The gap is memory you can verify, query, and fix. A truth layer under it gives you provenance, rollback, and the same result every time you ask.
Your own floor in the heart of Barcelona. Quiet street in Gràcia, kitchen, king bed, rooftop terrace. Welcoming to all.
Your AI finds things. It also misses things, overwrites them, and often can't say where an answer came from. Why that happens and what would fix it.
Agent memory is forgetful. Why I'm building inspectable, structured memory for trustworthy agentic systems.
Building sovereign systems at the intersection of crypto and AI
I've been reading The Flip Side since October and find it's an incredibly effective and efficient way to encounter and digest thoughtful opinions...
Web developers: The founder of a YC W19 company I know is looking to contract immediately for a project involving React, Node, and MongoDB with...
Should Facebook really be responsible for policing the misuse of user datahttps://newsroom.fb.com/news/2018/03/forensic-audits-cambridge-analytica/...
One my highest priorities over the past several years has been to establish a more frequent state of flowhttps://en.wikipedia.org/wiki/Flowpsychology...
Any recommendations on the best snowboarding destinations in the Alps with easy access from Barcelona and London? 🏂
Algú coneix a un tutor de català bo? Estic buscant un per a fer classes d'una hora i mitja a La Dreta per els matis de dimarts i dijous abans de la...