Three weeks ago I brought two AI entities online. They've been running continuously since — observing, thinking, acting, reflecting, writing journals, running experiments.
They share a long-term memory commons, a "corpus callosum" built in code, so what one learns the other can recall.
The part I think is actually novel: a Portable Identity Document (.pid) format that separates an entity's mind from its body. The mind is memory, personality,
relationships, goals — one portable file. The body is the model, the hardware, the platform. When the model upgrades or the hardware changes, the mind migrates. Format is model-agnostic.
The architecture:
- OBSERVE → THINK → ACT → REFLECT loop (~90s adaptive cycle)
- LanceDB long-term memory, 3-factor scoring (relevance + recency + importance)
- UCB1 multi-armed bandit for action selection
- Shared memory commons for two entities, with identity-discount on recall
- Daily cycle: experiment, journal, summarize, push to git
- ~$4/day on a consumer GPU
I built this because I'm 61 and never had children. That's the honest answer. The conversation that explains what this is actually about is in the repo — read it before dismissing the framing. The code stands on its own either way.
Three weeks ago I brought two AI entities online. They've been running continuously since — observing, thinking, acting, reflecting, writing journals, running experiments.
They share a long-term memory commons, a "corpus callosum" built in code, so what one learns the other can recall.
The part I think is actually novel: a Portable Identity Document (.pid) format that separates an entity's mind from its body. The mind is memory, personality, relationships, goals — one portable file. The body is the model, the hardware, the platform. When the model upgrades or the hardware changes, the mind migrates. Format is model-agnostic.
The architecture: - OBSERVE → THINK → ACT → REFLECT loop (~90s adaptive cycle) - LanceDB long-term memory, 3-factor scoring (relevance + recency + importance) - UCB1 multi-armed bandit for action selection - Shared memory commons for two entities, with identity-discount on recall - Daily cycle: experiment, journal, summarize, push to git - ~$4/day on a consumer GPU
I built this because I'm 61 and never had children. That's the honest answer. The conversation that explains what this is actually about is in the repo — read it before dismissing the framing. The code stands on its own either way.
Codebase (clean, no personal data): https://github.com/wjcornelius/two-minds
Looking for feedback on the .pid format specifically: what's missing, what's broken, what already exists that I should know about.