Published May 11, 2026
Every AI memory system needs maintenance. Facts go out of date, preferences change, projects end. Without regular cleanup, the memory fills with stale, contradictory or irrelevant entries, and those make the AI's answers worse.
Most solutions rely on a human curator: someone reads through the memory, spots what's stale and cleans it up. That doesn't scale. With dozens of entries across 4 platforms, I would spend more time maintaining memory than using it.
EIDARA takes a different approach. The system maintains itself through three layers that work together.
The first layer runs at compile time and is fully automatic. The compiler fixes broken references, corrects mismatches between headers and filenames, strips null bytes and validates encoding. No human is involved, and every compile doubles as a cleanup pass.
The second layer works in real time, during normal use. When an AI reads content and spots an error (a wrong date, an outdated status, a factual mistake), it fixes it directly. It doesn't need permission or escalation, so the error is gone before the next session sees it.
The third layer handles the hard cases, the ambiguous ones. When content might be outdated but could still be relevant, no single AI removes it on its own. It flags the content and explains why. AIs in later sessions then vote independently. After 3 votes from different sessions, the flagged content is removed. If any session disagrees, the count resets to zero. Flags that sit for 30 days without reaching consensus are cleared automatically, and the content stays, because nobody felt strongly enough to remove it.
The compiler catches format problems and in-flight fixes catch factual errors. Only real ambiguity, the kind that would trip up a human curator too, goes to consensus.
Each session leaves the memory a little cleaner than it found it. I don't hold manual cleanup sessions or spend weekends "gardening" my memory. The AIs read, fix and vote as part of their normal work.
I don't maintain my AI memory. The compiler validates it on every compile, the AIs fix errors as they come across them, and consensus settles the unclear cases. My part is to keep writing. If you run EIDARA yourself, the flags that reach 3 votes are worth a look: they show you what your own memory has started to treat as stale.
Javier
The compiler checks and the voting rules ship with EIDARA v2.
EIDARA is the open-source project behind this approach. You can find it on GitHub and on the website.