Tata

The Problem

Good financial advice is gated behind two things most people don't have: money and literacy. If you can afford an advisor, you get one. If you grew up fluent in budgets, index funds and savings rates, you can do it yourself. Everyone else is left with spreadsheets they never open and apps full of charts that ask to be loved. The people who would benefit most from advice are exactly the people least served by it.

The Solution

Tata is a personal financial advisor that lives in a chat you already have open. There's no dashboard, no forms, no jargon — you talk to her in plain words and she answers in plain words. You forward a bank statement, she reads it and sorts every line into its category. You tell her you want to retire by the sea, she remembers and bends her advice toward it. She runs entirely on your own machine, and she's free and open source — the whole point is that anyone, regardless of income or financial literacy, can have an advisor in their pocket.

One Agent Loop

Under the warmth is a deliberately small architecture: one Telegram bot, one Claude loop. The system prompt stays lean — persona, reply style, and an index of what she can learn — and everything else is pulled in on demand. Each turn, the agent first checks its memory, loads the one skill the task needs, calls a tool or two, and replies. No giant prompt, no sprawl.

Skills, Loaded on Demand

Instead of one bloated prompt that tries to know everything, Tata keeps her expertise in a library of small markdown skills — one per task. The schema for writing SQL, the recipe for parsing a bank PDF, the conventions for market lookups, the rules for editing memory: each is its own brief. The agent reads the user's intent, calls load_skill for the one it needs, and that know-how stays in context only for the turn that uses it.

Memory, in Two Stages

For an advisor, remembering is everything — and Tata remembers in two distinct layers, both per-user and both local. The first is the transcript: the last twenty turns of the live chat, kept in SQLite so the conversation feels continuous. The second is long-term memory: a small virtual filesystem under /memories that the agent reads and writes itself — your profile, goals, accounts, preferences and active worldview. The bridge between them is time: after six idle hours, Tata summarises the session into a single rolling note and starts the transcript fresh, so nothing important is lost and nothing stale piles up.

Tools & Database Access

Every number Tata quotes comes from a single SQLite file on your machine — every transaction, asset, loan and subscription in one place. The agent reaches it through a deliberately narrow set of tools, and the split between reading and writing is enforced, not just trusted. Questions go through run_sql on a connection opened query_only — it physically cannot mutate. Anything that changes your data flows through one path, apply_changes, which always shows you a preview and waits for a ✅ before a single row is touched.

Reading Your Statements

Getting data in is just as gentle. You export a PDF from your bank or card and send it as an attachment; a second ingest loop reads every line, sorts each transaction into the right category, and shows you a preview before anything is saved. If it keeps filing your favourite bakery under “restaurants,” you correct it once and a skip-rule remembers. Idempotent import hashing means re-sending the same statement never double-counts.

Why It's Open

Tata is MIT-licensed and runs on your own machine because financial advice shouldn't require trusting a company with your most private data — or paying for the privilege. The goal is simple and a little stubborn: a competent, patient financial advisor that anyone can install for free, no matter how little they earn or how little they care about money. You end up on top of your finances; nothing about them ever leaves the room.