Published May 17, 2026
Two months ago I shipped EIDARA v1.0, a constitution and a compiler that give any AI persistent memory across sessions. I've run it on my own work every day since. Today I'm shipping v2.0 with the lessons from that period, and describing where the project goes next.
v2.0 is the same EIDARA, now able to correct itself. These are the main changes:
The system asks you to maintain it instead of waiting silently. Before, when the Librarian (the hygiene pass) was overdue, only the maintainer noticed. Now every AI that reads BRAIN.md sees the status in the header and prompts you to run it. Missing neuron summaries, which inflate token costs without anyone noticing, get the same treatment.
The compiler checks that the Librarian really ran. A new anti-tampering check compares the Librarian's "last run" timestamp with the changelog. If the file says "ran today" but the changelog shows no Librarian activity for today, the compiler flags it as suspicious. Maintenance can no longer lapse without a warning.
The constitution now covers sandbox filesystems. Environments with FUSE mounts (dev containers, sandboxes, Cowork-style tools) can serve stale cached reads that look like file corruption. I learned this the hard way, and v2.0 turns it into a rule: never roll back based on a single read from inside a sandbox.
Smaller changes: session-closing summaries are capped at 500 words, large drops in neuron content are detected with a baseline check, and the Librarian's job is now spelled out, so format checking alone no longer counts as a full pass.
If you're running v1, the migration guide takes you through the upgrade in 5 minutes. Your VAULT data is preserved; only system files change.
v1 gave EIDARA memory and v2 made it self-correcting. v3 changes where that memory comes from.
I'm calling it SUPER DARA, and the idea is that your AI shouldn't need you to type everything it should know. It should read, listen, and see.
Reads your conversations and distills them into atomic facts stored locally. Before each new session, the compiler injects the most relevant context into BRAIN.md. You don't edit anything by hand, and the result is less drift, less effort and a lower token cost per session.
Listens to your meetings and captures what matters. Instead of a transcript dump, it keeps the specific decisions, action items and context that make your AI useful in the next conversation.
Sees what you've been doing on your screen. An AI agent starting a new session already knows about your recent work, so you don't have to bring it up to speed.
Everything is captured locally, queried locally and never sent to the cloud. Communication happens over MCP, the open protocol, so any compatible AI can plug in.
The design follows the same principles EIDARA has had since the first version: local-first, file-based, no infrastructure, fully auditable. It needs no new servers and no new accounts. As the state of the art improves, EIDARA takes the parts that work and keeps the constitution.
SUPER DARA is in the research phase. v2.0 is the stable release to use today, and the migration guide above is the place to start if you're on v1.
Javier
v2.0 is the stable release: self-checking memory you can install today.