Published May 30, 2026
I was the bottleneck. Every time my AI memory needed maintenance (fixing stale references, deduplicating entries, compressing bloated neurons) I had to be there. With 4 businesses and 38 memories to maintain, the system worked, but it depended on me opening a session and running the Librarian by hand. That doesn't scale.
I put a local LLM on the machine. It isn't a cloud API but a model running locally, 24/7, with one job: curate the memory autonomously. There is no human review loop and no approval queue. The constitution and git history are the controls.
The first iteration showed exactly where the gaps were. Within five hours, the autonomous curator had generated five false positives, flagging connections between unrelated projects because they shared generic tags. The model's JSON parse rate was around 48%. The system needed work.
I switched to a more capable local model, and the parse rate went to 100% overnight. I also tightened the tag matching threshold, added a blocklist for generic terms, and hardened the deduplication prompts. None of these changes was large, but together they removed the false positives.
One week later the count was 140 autonomous cycles, one every hour, with zero errors. The system was cross-referencing, deduplicating, and compressing without anyone watching. That was when I realized the architect role, the one I'd spent months building, had made itself nearly obsolete.
EIDARA v2 is open-source and free. What runs on my machine now goes further: a local LLM that curates your memory while you sleep. Many AIs contribute knowledge via MCP, and one autonomous curator decides what gets integrated. There is no cloud, no API key and no vendor involved, only a model on your own hardware keeping the context every AI session starts from up to date. The next thing to measure is whether that curator stays at zero errors once several AIs are writing to it at the same time.
Javier
A local curator that maintains your memory while you sleep. Early access is open.