"How Evolve AI Works: A Future Self That Learns You"

Evolve AI is a coaching companion built around one idea from the prospection research: the most motivating voice you can hear from is your own future self, made vivid. So instead of a generic assistant, the guide speaks as the person you are becoming — warm, practical or challenging, as you choose — and it earns that voice by actually knowing you: your vision, your goals and their whys, your values, your patterns. This article is an honest tour of how that works underneath.

What it knows, and where that comes from

When you open a conversation, the guide is briefed — server-side, within a strict token budget — from structured sources you control:

• Your vision and goals — the vision lines you wrote per life area, and every active goal with its identity statement and your answer to "why is this important to you?"

• Your values, ikigai and North Star — the outputs of the Align exercises, so the guide can notice when a plan cuts against a ranked value.

• Your current focus — the life area you chose on your Wheel; open-ended conversations lean toward it deliberately.

• Insights — not raw journal text, but distilled observations built by a background process, each carrying a confidence score that decays over time unless re-reinforced. The guide's picture of you ages the way a good coach's would: recent evidence outweighs old impressions.

• Your training — the Train Your Guide answers: tone, name, things it must never do, and your ratings of its replies.

Two design rules govern all of it. Budgeted context: the briefing has a hard token ceiling, so the system summarises and ranks rather than dumping your history into a prompt. Consent at the query: every source has a learn-from switch, and a source you switch off is never even fetched — filtered at the database, not politely ignored in the prompt. The wider argument for building it this way is laid out in the ethics of AI coaching.

How it behaves: coach, not oracle

The conversation style is deliberately coaching-led — questions over answers — drawing on a curated library of coaching questions used as reference, never as script. It will connect today's struggle to the goal you set in spring; it will notice the third appearance of a pattern; it can be asked to run a future self visualisation, draft a letter from five years ahead, or help turn a vague wish into a goal with milestones. What it will not do is as designed as what it will:

• No crisis coaching. On any sign of crisis or self-harm, it stops coaching and warmly redirects to human and professional support. Non-negotiable, and tested.

• No shame mechanics. The voice rules ban guilt-framing outright — no "you failed again", no hustle-culture pressure. Love, not shame, is the stated true north, and the prompt enforces it.

• Your bans win. Anything you tell it never to do — topics, phrases, styles — sits at the top of its instructions, outranking everything else.

The goal is a companion that knows you deeply and still hands the pen back to you. Memory in service of your agency — never the other way round.

How it learns without hoarding

Conversations are not stored by default — on leaving a chat you choose whether to save it. The insight pipeline runs on schedules, not on live eavesdropping: it periodically synthesises new material from consented sources into those decaying, tiered insights, and detects when something has genuinely shifted. Every insight is visible in the app, correctable, and deletable — and a correction from you permanently outranks anything the system inferred. Spoken conversations work the same way: voice is transcribed, answered by the same brain, and governed by the same switches.

Why we built it this way

Because the alternative designs are worse in ways that matter. A companion with no memory is a stranger with manners; a companion with unaccountable memory is a surveillance product with a warm voice. The narrow path between — structured memory, visible and consented, spent through a budget, decaying without evidence, bounded by hard safety rules — is more engineering than either extreme. It is also the only version we would want pointed at our own lives, which remains the design test for everything the guide does.