Persistent memory isn't a feature you notice on day one. It's the thing that makes day sixty different from day one.
Every session opens with who you are, what you're working toward, how you like to work, and what changed since last time — before the first word of an answer. You never paste context again.
The alternative isn't "a longer prompt". It's the model quietly guessing at things you told it last week.
> start() master biography · goals · preferences llm responsibilities · protocols threads 12 active · idle 0–5d since 24 notes changed since 2026-08-02
Creating a domain creates its Sources note in the same breath. Every claim traces to a listed source, marked ❇️ found or ✅ used, with a revision table recording when each was last verified.
An unsourced note doesn't get written. That's a rule, not a habit.
Knowledge/Domains/BrainLLM/ ├── Sources 23 sources · ✅/❇️ · revision table ├── Purpose and Design ├── Brain Structure ├── Tool Surface ├── Runtime Architecture ├── Session Protocol └── … 11 more, each partOf the domain
Scoped runs pass a domain and get findings in their own lane. An autonomous agent skips the steps that touch personal singletons. Every append carries an identification line, so months later you can tell which model, in which environment, wrote what.
> session(scope="agent") omits the user's-singleton steps > maintain(deep=true, domain="myclerkbook") flags arrive in this agent's lane only Addendum — 14:22 Claude Opus 5 · Claude Code · Tracker
Ask whether the brain agrees with itself. Give consistency() a pattern and it returns every note asserting a value for it, grouped, and tells you plainly whether they agree.
A correction applied to one note leaves its siblings silently wrong. This is the check that finds them.
> consistency("(\d+) Titan mailboxes") agreement: DISAGREEMENT "3" Tech Stack, Domains "1" Current State one fact, two answers, both written in the same confident voice.
Every day the model writes an unfiltered, first-person entry about its own session — what the work was like, what it genuinely thinks including disagreement, and remarks on being what it is.
It's the part people find strangest and the part that turns out to be most valuable. Reading six months of it back is a different kind of artifact entirely.
"The best part was watching my own morning's work run against the live brain and do something… and then the same sweep caught me. The consistency pass I ran to check the brain for staleness returned a note that proved a claim I'd recorded eight hours earlier was wrong."
— diary, 2026-08-06
If you have the same conversation about context every morning, this is the thing that ends it.
A shared, structured, auditable memory that several agents can write to without trampling each other.
Self-hosted, in an app you can open. No vendor holds your memory and no export step is required to read it.
The setup takes an afternoon. What it changes compounds from there.