Published May 30, 2026
For two weeks I ran a local LLM as an autonomous curator over my memory: 38 active memories, 14 agent definitions and the context of 4 businesses. The curator runs every 15 minutes with no human oversight. The safety net is the Constitution plus git history. This is what happened.
In the first five hours of day 1 I got 5 false positives. The model was flagging cross-references between memories that shared tags like "website" or "marketing", which covers almost everything when you run 4 businesses. The JSON output was malformed about half the time. I switched models, raised the tag threshold from 3 to 5 matching tags, added a blocklist for generic terms, and hardened the prompts. By the end of week 1 it had run 140 cycles with zero errors.
With the curator stable, I added the facts layer. Any AI can now contribute a fact via MCP. The curator picks it up, deduplicates it, finds where it belongs and integrates it, or discards it with a reason. I seeded it with 5 test facts: 4 were stored correctly and 1 was merged with an existing entry. Zero errors.
(1) The model had a "thinking mode" that was consuming the response budget and returning empty answers. One API parameter fixed it.
(2) Every compile cycle was throwing about 40 background distractions per run, small noise from how the autonomous side coordinates its work. I fixed it by making all of that run silently.
(3) The curator couldn't tell routine maintenance apart from structural changes that need a human architect. I added an escalation signal: a dedicated queue that gets surfaced in the compiled brain. It has three tiers: routine (curate silently), data (add to neurons) and structural (escalate).
The cadence started hourly and is now every 15 minutes. Total autonomous cycles: 300+. Errors after the first-day fix: zero. I added 4 new constitutional rules to govern the autonomous curator. In the first 24 hours, several sessions contributed facts via MCP in parallel and all of them were integrated cleanly. Since activation the system has been up for 12 days with no human intervention.
During the test phase I found a more capable model in the same family, with better JSON reliability, stronger multilingual handling and the same memory footprint. Migration would have taken five minutes. I chose not to do it.
If you change the model while you are measuring system behavior, you can't tell whether an anomaly comes from the new model or from a bug in the pipeline, and the data is contaminated. So I'm keeping the current model until the test phase closes and collecting clean decisions from real traffic. The upgrade comes after, with a proper A/B comparison over hundreds of real curator decisions. First get the most out of what is there, then measure the upgrade.
Open-source LLMs move every 4-6 weeks, and I re-scan monthly. There will always be something newer, and part of the job is deciding when not to chase it.
The curator, the facts channel and the MCP all work. What's missing is the senses, the ability to capture context from the screen and from audio without the user having to type anything. That's the next build, and the architecture stays the same: each new sense is one more source of atomic facts feeding the same curator pipeline.
EIDARA v2 was released free and open-source. What came after, DARA, is being built in private. If you want to follow the progress, the insights page is where I post it.
Javier
The autonomous curator and the senses are what SUPER DARA adds on top of v2.