What’s changed

The free trial is gone, in the good direction. Refractive includes open models with no message limit and no API key required. Connecting your own key is now an upgrade you can choose, not a wall you hit.

Every request to the model service carries a requirement not to store or train on what is sent — and a refusal to be re-routed to a provider that would. If no provider can meet that, the request fails rather than quietly going somewhere it shouldn't.

The privacy policy named Anthropic and ElevenLabs as the recipients of your conversations and your agent's voice. Since August 16 those were OpenRouter and Cartesia. Both documents now name the right companies, and take their names from the same place the code does so they cannot drift again.

The terms said you must bring your own API key after a trial. You do not. They also now say plainly that Refractive intends to charge for itself eventually, rather than leaving 'free' to be read as a promise.

Export my data — one file containing every record tied to your account, from conversations and context entries to each life domain's history. No request form, no waiting.

Delete my account — removes your sign-in and everything attached to it, immediately and permanently. Before you confirm, it shows you exactly how much will be erased.

A privacy policy and terms of service written to be read, including a plain account of where your words actually go when the agent answers you.

If you share a Finances household with someone, your leaving no longer takes their records with you. The household stays, with every transaction and budget intact, and ownership passes to the longest-standing remaining member.

This page. Every meaningful change to Refractive, newest first, written for people who use it rather than people who build it.

The changelog is listed in the site map and in llms.txt, so AI assistants answering questions about Refractive can see the current state of the product rather than a snapshot from whenever they were trained.

A schema health check that verifies all 44 data tables and 18 specific fields Refractive depends on, and reports in plain English which feature would break if one were missing.

All database setup consolidated into a single operation that is safe to run repeatedly — no more partial or ambiguous upgrade state between releases.

Repaired a broken toolchain install that had been masking type errors for several weeks, then re-reviewed the highest-risk recent changes — trial-message accounting, sign-in, and encrypted key storage — against a clean baseline.

Retired an old, unused conversation endpoint that bypassed the bring-your-own-key routing. Nothing called it; leaving it around was a liability, not a feature.

Every AI task in the product is now classified by complexity. Routine background work — tagging, extraction, classification — routes to an efficient open-source model, while live conversation always runs on the best model you've connected. There is nothing to configure.

A read-only status panel in Settings showing exactly how your work is being routed, because the reasoning should be visible even when the mechanism is invisible.

Two more reference guides: how to organize context files for AI, and how AI life-management agents compare by category.

The Context Graph dims deactivated domains and labels them inactive rather than hiding them. Deactivating a domain never deletes its history, and the graph is meant to show everything the agent knows — hiding a domain there would have contradicted both.

Shared households in Finances. Invite another person by link, see the roster and pending invites, and track a budget together. Invites expire after 14 days.

Choose which life domains you actually want tracked. Deactivating a domain removes it from your map and stops the agent filing anything there; reactivating it restores the full history untouched.

Custom agent modes. Save your current agent configuration under a name and switch between named modes instantly, alongside the six built-in ones.

Controls for how sessions close, whether the agent surfaces things proactively, and how often it asks clarifying questions.

Site-wide groundwork for AI-assistant discovery: structured data, a machine-readable product summary, a site map, and explicit permission for AI crawlers.

Vocabulary level, profanity, metaphor use, and cultural reference palette are now actually applied to every conversation. Previously they saved correctly and were then ignored.

A cultural reference palette — pick up to three domains the agent is allowed to draw analogies from, so its metaphors land in a world you recognize.

Five avatar styles, five color themes, and four animation personalities, with a live preview while you choose. Some unlock as you use the product.

Bring your own 3D model — point the agent figure at your own file if you'd rather.

Your YOU node's completeness always read as zero regardless of how much the agent had learned about you, which also meant one of the four unlock milestones could never be reached. It now reflects what you've actually shared.

Four milestones now unlock progressively: seven sessions, five completed missions, thirty days with an account, and an 80%-complete YOU node.

Relationships: your inner circle, who you've drifted from, upcoming birthdays, and one observation from the agent each time you visit. Deliberately not a CRM — there is no contact database to maintain here.

If you use Social Cycle, Relationships reads from it directly. If you don't, the agent learns your inner circle through conversation instead.

Career & Mission: opportunity pipeline, conversation prep, and a mission statement that feeds back into your YOU node. The agent pushes back here rather than agreeing — it will not validate weak reasoning, and it will not make the decision for you.

A data bridge to Social Cycle so professional contacts can inform career conversations, with a clear boundary: Refractive reads contacts and writes only one field back.

Contribution: a values interview, four pillars of giving, commitments the agent tracks, and a legacy statement that enriches your YOU node. Strictly non-partisan by design.

Spirituality: practice accountability without streaks or guilt, a reading list, and an agent that challenges your practice while never taking a position on your beliefs.

Finances: a six-phase budget interview, a Finance Score, and statement import — upload a CSV export from your bank or card and it's parsed, de-duplicated, and categorized automatically at roughly 83% accuracy on the first pass, improving as it learns your merchants.

Partial months are flagged before any total is quoted, so a mid-month import never quietly reports a misleading number.

Session-based reflection with a longitudinal model that remembers across sessions, plus private mode for conversations that are never retained, and per-entry hide and delete.

Crisis guardrails that direct people to real human support. This is a reflection tool, not therapy and not a medical device.

Physical Body: a Vitality Score, a health interview conducted by voice, and weekly check-ins.

Time: read-only Google Calendar sync, habit baselining, task classification, and a time-block view. Time acts as the orchestration layer across the other domains.

The agent figure's mouth now actually moves in time with speech. Two separate causes: the 3D model had no jaw to animate, so one is now derived at load time; and routing playback through the audio analyser could mute the agent entirely if the browser had suspended audio.

A fourteen-attribute character creator — warmth, directness, formality, humor and more, each on a nine-point scale — set before the first interview and adjustable at any time.

Six built-in agent modes, plus a My Agent settings area.

The Life Graph engine — conversations are classified into life domains and stored as linked context entries the agent can reason across.

A library of 174 project archetypes the agent draws on when suggesting what you might take on next.

refractive.tech is live.

Connect your own Anthropic, OpenAI, or Google API key and pay the provider directly. Keys are encrypted at rest, and you can remove one at any time.

A free trial on Refractive's own key, so nobody is stopped at the door by a developer signup.

Sign in with Google, alongside email and password. Worth being explicit that this is unrelated to which model runs your agent — no provider lets a consumer account grant a third-party app inference access.

Resolved a bug where starting a voice conversation did nothing at all.