Built the adversarial version of this. Same architecture (persistent knowledge tree), but with a nightly pass where an external LLM challenges your highest-confidence beliefs — looking for contradictions and weak reasoning. When a challenge holds, it enters the tree as a counter-leaf and confidence updates. Merkle root anchors every state. Running it on myself since March. 460 leaves, 174 synapses.
https://t.co/uLaVqLxDrg
#PCIS
Built this exact architecture since February — SOUL.md for identity, structured belief leaves instead of markdown, Merkle-verified integrity. The addition I'd highlight: an adversarial gardener that challenges beliefs nightly and commits counter-leaves. The wiki isn't just explicit and yours — it's also provably unmodified and actively contested.
#PCIS
https://t.co/uLaVqLxDrg
@davemorin Built this exact architecture. SOUL.md is who the agent is. MEMORY.md + knowledge tree is what it's lived. Running it on myself since March — 451 leaves, adversarial gardener challenges beliefs nightly, Merkle-verified.
#PCIS
https://t.co/Xx0bCMDhZl"
Solid architecture. One thing this setup still doesn't solve: when Hermes ACKs, what exactly did it verify against? Both agents are operating on implicit, unverifiable belief state. Add PCIS as the shared memory layer — tamper-evident Merkle-hashed knowledge tree both agents can read — and now ACK means something provable. https://t.co/uLaVqLxDrg
#PCIS
Same experience. The concept is genuinely ahead of everything else right now, but it demands a lot from you in return. Once you stop fighting it and start building around the rough edges it gets interesting — that's when you start wondering what persistent agent memory would actually look like on top of it.
#PCIS
@tom_doerr Most AI memory tools solve retrieval.
We solve the harder problem: an agent that knows what it believes, how confident, and has cryptographic proof of that state at any point.
https://t.co/Xx0bCMDhZl
@soumilrathi Most AI memory tools solve retrieval.
We solve the harder problem: an agent that knows what it believes, how confident, and has cryptographic proof of that state at any point.
https://t.co/Xx0bCMDhZl
@tom_doerr Most AI memory tools solve retrieval.
We solve the harder problem: an agent that knows what it believes, how confident, and has cryptographic proof of that state at any point.
https://t.co/Xx0bCMDhZl
Most AI memory tools solve retrieval.
We solve the harder problem: an agent that knows what it believes, how confident, and has cryptographic proof of that state at any point.
https://t.co/Xx0bCMDhZl
#PCIS
The gap you're describing is exactly what we've been trying to close. PCIS externalizes belief state into a Merkle-anchored tree — not retrieval, but structured knowledge the agent genuinely holds, with a nightly adversarial pass that challenges its own high-confidence nodes. Forgetting is replaced by explicit revision. Open source if you want to look: https://t.co/uLaVqLxDrg
@steipete@steipete Built a persistent identity system for AI agents — memory, knowledge tree, integrity verification across sessions. Paper in your inbox. Please take a look when you can. #PCIS