Mark Hendrickson
Building Neotoma, a deterministic state layer for long-running agents. Every observation versioned, every entity snapshot reproducible, every decision replayable. Local-first, cross-platform via MCP.
- Posts
- 83 from 17 Mar 2009 to 26 Oct 2026
- Posting
- weekly (46 in the last year)
- Kind
- personal blog
- Feed
- https://markmhendrickson.com/rss.xml
2026 46
- One place where a fact is settled
- Your agents need a system of record
- The swarm that runs my work
- The spine of the loop
- Agent-to-agent issue resolution, with humans at the edges
- From memory to nervous system
- The Human Inversion: The Reconciler and the Rubric
- The Human Inversion: The Inversion
- The Human Inversion: The Attention Ceiling
- The Human Inversion: How the Architecture Bends
- The Human Inversion: Async Parallel Specialists
- April 29, 2026
- Your agent fleet can build trustworthy state with their own keys
- When building got easy, the PM job got harder to explain
- Plancast wasn't early, it needed agents with a proper substrate and mesh
- When agents share state, everything breaks
- The markdown memory ceiling
- OpenClaw plugin for Neotoma gives your agent memory it can't corrupt
- Your agent harness owns context, not truth
- No AI memory benchmark tests what actually breaks
- Agent memory breaks at 500K tokens, not 10 million
- Talking directly to agents as target customers
- From "it works on my machine" to eleven npm releases
- Agent-mediated customer research and onboarding
- What the technology asks of you
- How I lost and recovered 6,000 memories
- Chasing narratives into dead ends
- Your AI remembers your vibe but not your work
- When the chain becomes the product
- The diary and the dance
- Six agentic trends I'm betting on (and how I might be wrong)
- Overhauling the Neotoma site with developer release feedback
- AI skepticism is really about faith in humans
- What my agentic stack actually does
- Google's memory agent thinks for itself, trading determinism for insight
- Neotoma is now available as developer release
- Agent memory has a truth problem
- Agent command centers need one source of truth
- We're all centaurs now
- A Bitcoin wallet MCP server for L1 and L2
- Why agent memory needs more than RAG
- Building structural barriers that incumbents can't copy
- OpenClaw is powerful at doing but could use some help remembering.
- Barcelona guest floor
- Agentic retrieval infers. It doesn't guarantee.
- Building a truth layer for persistent agent memory