Published May 25, 2026
A week ago I turned on the autonomous side of DARA. This is the update, with the numbers.
It looks after 38 memories and 14 working agents, and over the past week it ran about 140 times, once an hour for every hour the machine was on. Each cycle follows the same routine: scan every memory, write a one-line summary for any that lacks one, look for memories that should be cross-linked, clear dead references, flag what’s old enough to archive, and rebuild the compiled context the system reads from. That context is roughly 7,500 tokens, distilled from about 97,000 tokens of raw memory. That’s a compression of about 13×, recomputed on every cycle. This is what agent memory consolidation looks like in practice: persistent memory that stays useful without human curation. Across the whole week the log shows zero errors.
Two things changed along the way, and I want to be open about both.
The autonomous side runs on a model that lives entirely on my machine, so nothing leaves the device. That’s the local-first principle EIDARA is built on. The first model was a little too simple for the judgment calls the work needs, so I moved to a more capable one. The field has a name for this kind of component now: sleep-time compute, an agent that does its thinking in the downtime between sessions. That describes DARA’s autonomous side well. It is a sleep-time agent that curates and consolidates memory while you’re away.
A diagnostic at the end of the week found a gap. The system was good at the routine work, but it couldn’t tell the difference between “something I can fix myself” and “something only a human should decide.” So it did neither, and those items just piled up. I built it an escalation signal: now, when the autonomous side hits something structural, it raises a flag that the next session sees immediately. That’s the loop I keep repeating: find where the judgment falls short, fix it, and let it run again.
The bigger shift is in how I think about what comes next. I’d been treating DARA’s roadmap as a list of features: memory, then eyes, then ears. I now think that’s the wrong frame, and that a body with senses is a better one.
There will be one common channel, a facts layer, where everything DARA perceives lands as small, atomic facts. The autonomous side curates that channel. It decides what’s new, what’s already known, and what contradicts something it already holds. Each sense is then one more input into the same channel. Reading is one. A memory connection through which other AI tools can contribute (via something like MCP) is another. Screen and audio come later, into the same channel with the same curator. This is context engineering at its most literal: designing what context reaches the model and how.
The first step is the facts channel itself, with reading and the MCP connection as its first two inputs. After that I’ll report whether the curator keeps up once several senses feed it at the same time.
Javier
A memory that tends itself between sessions. Early access is open.