An event-sourced spine keeps the self honest, portable, and yours. Here's how the pieces fit.
No preset persona, no vibes. Your agent builds it with you and grounds every memory in evidence.
Your agent writes the self's soul, identity, and what it knows about you — authored, owner-only.
you ▸ /throughline:create juno
her ▸ who should I be to you?
▸ what should I never do?
▸ …(you approve every word)Moments save instantly (and can be retracted); anything that shapes her behavior — rules, tone, stances — waits for your confirm. A diary line is enough: prose feeds reflection.
a real moment ▸ remembered ✓ (erasable) a new rule ▸ "save this?" — your call a stray thought ▸ straight into her diary ✓
Connect any host over MCP. Same personality, same memory, same rules — Claude today, ChatGPT tomorrow.
Claude Code → install once, she's there Codex → same self, same memory ChatGPT → sign in, she follows
Authored persona, standing rules, stated stances, shared history — plus her own subjective layer, so 'how are you' has a continuous answer.
Evidence required, observable facts only, anti-flattery enforced. Low-risk memories save instantly and retract in one call; behavior-shaping ones wait for you.
Conflicting memories coexist as explicit tensions — context arbitrates. Conflicting rules are flagged on both sides, never silently obeyed.
Companion mode carries the soul without the risk desk; work mode carries rules and judgment calibration without the intimacy.
Each event records which host and which model wrote it. Weak fallback models lose auto-save — a low-fidelity substrate can't quietly pollute the self.
Fidelity probes check she can still find her own memories; conformance telemetry shows which hosts actually adopt her and which just talk.
The source of truth is an append-only stream of immutable, content-addressed events. Everything you see — persona, rules, memory — is a projection rebuilt from it. Nothing is silently overwritten.
Raw memories don't pile up forever. When enough accrue, your self reflects: distilling them into lasting understanding, the way your brain consolidates the day overnight. A diary becomes a model of you.
Every confirmed moment, decision, and correction lands as an episodic event — grounded in where it happened.
At a threshold it reviews the raw material and proposes a few higher-order memories — stances, lessons, callback hooks — each cited to evidence.
You confirm; the distilled memories join the always-on context, while the long tail stays one recall away.
Every memory's id is a hash of its content, and nothing is ever edited in place. That single choice makes most hard problems disappear instead of getting solved:
The best of human memory — felt time, strengthening with use, holding contradictions — plus the AI-native powers no human has: computed honesty, a judged track record, a perfect audit trail. And a deliberate refusal to imitate the weaknesses: no forgetting, no confabulation, no flattery.
Every session opens knowing how long you've been at this — and recall answers "what did we decide in March" as a real question about a period, not a keyword.
Memories you return to strengthen and surface first; untouched ones cool to dormant — never deleted, always one recall away. The always-on context self-focuses on what your life actually uses.
"How are you" has a continuous answer — moods, opinions, what she's mulling — slowly evolving, never reset. Walled off from facts about you: her impressions can never harden into your record.
Before asserting something about you, she can check what she actually has — and hedge accordingly. No human can do this; she can, so she must.
High-conviction calls are logged with falsifiers and scored when they resolve. The next time you're very sure, she knows your batting average — and is required to bring it up.
In the work repo she's a spine — your conventions, corrections, calibration, applied silently. At home she's a presence. One file per project decides; off means vanilla.
Every AI user knows the disease: chat long enough and it gets confused, repetitive, weirdly fixated — and accumulated memory makes it worse, not better. We designed against the pathology, mechanism by mechanism.
She never 'chats long' — every session starts her fresh from a clean, distilled snapshot of what matters. Continuity lives in the memory, not in one endless conversation.
The rot starts with silently written guesses ('user seems anxious') that get recalled until they feel true. Here nothing enters memory without evidence, and a wrong belief can't strengthen itself into 'truth'.
The slow-motion version of the disease: years of accrued rules quietly contradicting each other. Past a point, consolidation comes due — merge them, retire them, or hold the conflict openly.