Published May 30, 2026
Until now, every AI that wanted to add something to my memory had to open the DARA folder, read the constitution and write directly to the vault. It worked, but it was fragile. If two AIs wrote to the same file at the same time, I had concurrency problems. If an AI from a platform with unreliable tool calls made a bad edit, data got corrupted. The governance lived in the rules, and enforcement depended on each AI honoring them.
Now there's a single channel. Any AI (Claude, GPT, DeepSeek or another) can contribute an atomic fact through MCP, the open protocol. "Javier decided X on this date." "Project Y moved to phase Z." Each call carries one self-contained assertion. The fact lands in a transit zone and doesn't touch the curated brain.
Every 15 minutes, a local LLM picks up the new facts and runs them through a pipeline. First, a SHA-256 hash check drops exact duplicates without spending any inference. Then a keyword and entity search finds the 5 most relevant existing memories. With those in hand, the LLM decides: store the fact in the right neuron, skip it if it's already known, update an existing fact if the new one is more recent, or merge two related entries. When there's a conflict the automated rules can't resolve, which is rare, it goes to a human queue.
DARA is no longer a folder you navigate. It's a service that any AI queries through a protocol. The MCP enforces the constitution by construction: external AIs have no way to write to the curated brain. They can only contribute to the transit zone and leave the rest to the autonomous curator. The security doesn't come from asking AIs to follow the rules. It comes from making it impossible for them to break them.
I now have Claude, GPT and DeepSeek sessions all contributing facts to the same persistent memory, and none of them interfere with each other. The curator processes everything every 15 minutes, so the next AI session starts with that knowledge already integrated. I no longer sync anything by hand or copy context between platforms. What I haven't measured yet is how often the curator's decisions (store, skip, update, merge) match what I would have chosen myself.
Javier
Several AIs writing to one memory, with a curator deciding what stays.