Ari: the AI accountability assistant you text

Ari, the AI accountability assistant you text, manages your life, frees your time, and makes sure you reach your goal. Send a message on iMessage, Telegram, or Discord, or use the web chat at textari.chat, and Ari handles the rest. Ari is artificial intelligence, not a human, and not a therapist or a crisis service.

I am Michael Saul. I build Ari at Saulutions. Why I built it is in Why I built Ari.

Start here: textari.chat · Pricing · FAQ · Roadmap · Guides and feature notes

What Ari does

Need	What Ari does
Habits	Tracks habits, logs done and missed, keeps streaks visible, and adjusts the plan when a streak breaks
Reminders	Sends one-off and recurring reminders on the channel where you already reply
Follow-up	Follows up on goals, deadlines, and commitments until they are done
Memory	Remembers goals, people, routines, deadlines, and preferences, and shows what is stored so you can correct it
Life admin	Works with calendar, email, tasks, notes, and files after you connect those tools
Research	Searches current public information, reads pages and documents, and pulls YouTube transcripts
Live browser	Checks stock, prices, and hours on the real site, fills out forms, and checks order status. You can watch live and take over. Ari asks before anything that spends money or submits
Ari’s own computer	A private cloud computer for files of any type (PDF, Office, photos, video), installing tools, running commands, and building things. Your files show in a Library. Read more in Ari can do what you’d sit down at a computer to do
Small apps and sites	Builds small web apps and simple websites on request, each with its own shareable address
Group planning	Helps in shared chats without loading private DM memory into the group

Channels

• iMessage: text Ari like any contact.
• Telegram: message the Ari bot.
• Discord: message the Ari bot.
• Web chat at textari.chat: the same conversation, once you have an account.

One conversation follows you across channels. Your first message creates the account; there is no form to fill out.

Architecture

The hard part is deciding what context and tools each job may use.

The main loop does roughly this each turn:

1. Normalize the message from chat, the website, or a scheduled wake-up.
2. Identify the user, channel, conversation, permissions, and job type.
3. Load only the context that is relevant to that job.
4. Decide whether to answer directly or route a bounded task to a scoped worker.
5. Run allowed tools and verify tool results before claiming work was done.
6. Record an operational trace for reliability debugging.
7. Clean up the turn workspace.

A good assistant needs memory and tools. Each message still gets only the permissions that job requires.

Memory and profile

Ari has several memory layers instead of one giant blob.

Layer	Purpose
Short-term history	Recent conversation context
Long-term memory	Durable facts retrieved by relevance
Structured profile	Name, timezone, goals, habits, preferences, important dates, people, and current context
Profile notes	Voice, boundaries, identity, and style preferences
Review pipeline	Background review of older messages for durable things worth saving

Memory is useful and correctable. You can ask what Ari knows, ask Ari to forget something, change preferences, pause check-ins, disconnect integrations, export your data, or delete the account.

Permissions and sandboxing

Integrations are opt-in and you can disconnect them at any time. Ari explains when an integration was used. Group chats have a smaller tool surface and never receive the owner’s private memories or DM transcripts.

The security model assumes the model can fail. A webpage, email, or PDF may contain instructions aimed at the model, so the defence is structural: each worker receives only the access its job needs.

• Public-web research gets public-web tools, never private memories.
• Integration work gets the relevant provider tools, never the full profile and chat history.
• Scratch workspaces are temporary and wiped after the turn.
• Code helpers run with restricted filesystem, network, and resource access.
• Credentials stay outside generated code and flow through controlled integration paths.

Ari confirms provider actions through tool results before reporting that they happened.

Scoped workers

This is where Ari is different from a generic “one big agent with tools” setup.

Ari has a few worker types, and each sees different things:

Worker	What it does	What it cannot see by default
Internal recall	Searches Ari’s user-scoped memory and app state	Public-web browsing and connected external accounts
Integration	Works through a specific connected API for a specific task	General internet access, private memory, unrelated provider scopes, and app internals
Web	Searches public information, reads public pages, and handles public media work	Private user state, connected accounts, internal app state, private files, and internal URLs
Reviewer	Looks for durable facts worth saving as short-term context rolls forward	User-facing chat, arbitrary tools, and external integrations
Group chat	Helps with shared planning and shared context	Private DM memory and owner-only context

The split is an access-control boundary.

Provider actions are confirmed through tool results before Ari claims they happened. If a task needs an integration, it goes through the integration lane. If a task needs public research, it goes through the web lane. If a task needs private recall, it stays inside Ari’s own data boundary.

Web vs integrations vs internal

The three main lanes are intentionally separated:

Lane	Can access	Cannot access
Web	Public web search, public page fetches, public documents, YouTube transcripts, public image search, and public media handling	User database state, private memories, profile notes, chat history, connected integration accounts, credentials, internal URLs, and app internals
Integration	The specific connected provider API needed for the job	General internet access, private memory, profile notes, full chat history, unrelated provider scopes, and app internals
Internal	Ari’s own user-scoped data: memories, profile, habits, important dates, observations, recent messages, schedules, and app state	Public internet access, arbitrary web fetches, connected integration accounts, and provider APIs

This is the part I care about most. The web lane is sterile. It can look things up, but it does not know anything private about the user and it cannot reach their integration accounts. The integration lane can talk to the specific APIs the user connected, but it cannot browse the general internet. The internal lane can answer private recall questions, but it cannot access the internet or external accounts.

If a prompt injection lands in a webpage, it is contained in the web lane. If it lands in an email or calendar event, it is contained in the integration lane. If the model gets confused while doing internal recall, it still does not suddenly gain web or provider access.

Proactive follow-up

Ari can send a one-off reminder, run recurring pings, check in on a habit, and follow up on a deadline or commitment. Check-ins are bounded: you can pause or change them whenever you want.

Limits

Ari is early and sometimes wrong. A follow-up can land at the wrong time, a tool call can fail, and the model can state something with false confidence. Ari does not give medical, legal, or financial advice.

Links

• Product: textari.chat
• Pricing: textari.chat/pricing
• Questions: textari.chat/faq
• Build story: Why I built Ari
• Maker: Michael Saul on GitHub