Living memory for a local AI.
A memory system that gives a local model a continuous life. Every conversation is written back to a durable ledger, indexed by meaning, and reassembled into context on every reply; identity, relevant memories, and the living conversation, resurrected each turn. It has run in my house for months. 111,000 memories and counting.

The model is stateless; the memory is not. Each turn is rebuilt from four layers; a fixed identity, a memory scroll pulled by hybrid full-text and semantic retrieval, the recent conversation, and a capsule each instance leaves for the next. The model can also reach back itself, with search tools answered straight from the archive. All of it runs locally; the model is on my own machine, and nothing leaves the house.
Running privately, continuously, for months. The version anyone could use; import your ChatGPT or Claude export and keep your own history alive on your own hardware; is where this is headed. It isn't packaged yet, and I won't pretend otherwise.
Devlog
live from git- Just shipped; the first entries are on their way.
If you want an AI that actually remembers, your team's, your family's, your own, this is the architecture I'd build it with.

