Published May 11, 2026
When you design a system, the natural instinct is to control who can do what. Only the authorized agent writes to this domain; only the admin modifies that configuration. It looks clean, secure and controlled.
I've seen this pattern play out in every company I've worked with. You restrict access to protect quality, but what you actually create is friction. The person with the information isn't the person authorized to enter it. The one who spots the error can't fix it directly. The right knowledge keeps surfacing in the wrong context, and the system's own rules prevent it from being captured.
Apply this logic to AI memory and the same thing happens. New information surfaces in a DeepSeek conversation, but that model "isn't authorized" to save it. Claude spots a factual error, but it's not the "designated writer" for that domain. The right information keeps ending up in the wrong session, and nothing gets saved.
Access restrictions create bottlenecks when they're about WHO rather than HOW. You lose information, you wait, you switch context, and you burn tokens re-explaining things the system should already know.
An efficient organization doesn't say "only these 3 people can access the system." It says "to access the system, you do it this way," and puts the controls on the process instead of on the people.
So let anyone write, and let the compiler enforce quality. Rules control, not roles.
EIDARA has a 15-rule writing protocol (W1 to W15) that governs every interaction. Any AI that follows the rules can read, write and fix errors. There are no special permissions per agent.
If the information is wrong, the AI fixes it directly, without asking for permission. If it's unsure about some content, it flags it with a consensus vote and other sessions weigh in. If the issue is structural, it leaves feedback in INBOX, and the system processes it on the next maintenance cycle.
The access control is the protocol itself, not the user or the platform.
Any AI can contribute at any time. Claude writes context in Cowork, DeepSeek writes context in TypingMind, and GPT writes context when needed. All of them follow the same rules, and everything they produce gets validated, deduplicated and compiled.
Nobody waits for authorization, because there is nothing to authorize. If you use more than one model, the test is easy to run: give them the same written rules, let all of them write, and check what the next maintenance cycle has to correct.
Javier
The 15 writing rules ship with EIDARA v2. Any model that follows them can write to your memory.
EIDARA is the open-source project behind this approach. Code on GitHub, more on the website.