Story How it works Tests Insights Whitepaper What's next Download Author Contact

Where EIDARA is going

Three versions. One principle:
your data stays on your machine.

v1

Memory

Shipped

A constitution and a compiler that give any AI persistent memory across sessions. Write what matters in markdown, compile it to one file, and every AI reads the same truth.

Neurons

Each project, person, or idea lives in its own markdown file. Atomic units of knowledge.

Compiler

One Python script reads your vault and produces BRAIN.md — the single file your AI needs.

Constitution

15 rules that govern how any AI interacts with your data. Model-agnostic, always enforced.

v2

Self-correcting

Shipped

The system maintains itself. Ignore it for weeks — it pulls you back before anything drifts too far.

Librarian self-trigger

Every AI reading BRAIN.md sees if maintenance is overdue. Alert threshold: 3 days.

Integrity checks

Protected file sizes tracked. Any unexpected drop fires a CRITICAL alert with rollback pointer.

Anti-tampering

Cross-references timestamps against changelog entries. Suspicious activity gets flagged automatically.

v3

SUPER DARA

In progress

Your AI shouldn't need you to type everything it should know.
Three intentional senses — you decide what enters, every time.

Voice

Push-to-talk dictation. You hold a key, speak, and only what you intend gets captured — distilled into atomic facts stored locally and compiled into BRAIN.md.

Typing

Only what you confirm with Enter. Nothing is captured as you type — just the decisions and context you choose to commit, so your AI is useful next time.

Screen

Only when you dwell on a window for more than 10 seconds — and even then, only the text the app exposes via standard accessibility APIs, never pixels. A new session arrives aware of your recent work, no briefing required.

Sensitivity modes let you mute any combination in one click.

Built on MCP — the open protocol

Any compatible AI can plug in. No vendor lock-in. Model Context Protocol is the standard — EIDARA speaks it natively.

v3 step 1 — running — internal testing

v3, one week in.

The autonomous side of DARA has now run for weeks in private testing — hundreds of cycles, quietly curating the 38 memories and 14 agents it manages, with no human in the loop and zero errors. Its cadence has tightened over time: it began curating once an hour, then every 15 minutes, and now every four minutes — so new context gets absorbed almost as fast as it arrives. We upgraded its local model to a more capable one, and taught it to raise a flag when it hits something only a human architect should decide. What comes next is a single channel — a facts layer — that turns everything DARA reads, hears and sees into atomic memory the system curates itself. A body with senses, built one sense at a time. Public release when it's earned. Read the full notes →

v3 step 4 — senses shipped — now stabilizing

v3, three senses in — and hardening.

The memory perceives in three intentional modes: voice (push-to-talk dictation), typing (a commit-aware keyboard hook that learns only what you confirm with Enter), and screen text (read through the Windows accessibility API — not computer vision). A confidentiality selector in the system tray pauses everything in one click, with a never-tunable privacy floor for password managers, banking, and sign-in pages. The pilot is done and the senses are live; the work now is hardening — replacing each rough edge with something that won’t break the same way twice. Read the latest update →

v3 step 2 — facts layer + memory access — designing

One common channel for everything DARA perceives. MCP as the input protocol. Each sense as a plug-in.

Same principles. Always.

No new servers. No new accounts. The state of the art keeps getting better — EIDARA absorbs the good parts and keeps the constitution.

Local-first

Your files, your machine

File-based

Markdown all the way down

No infrastructure

No Docker, no API keys

Fully auditable

Every byte is readable text