wikiTaTa adds lasting project memory, a vault your AI can use but never read, and automatic safety checks to the assistant you already use.
AI assistants are brilliant inside one conversation and forget everything by the next. They also ask you to paste passwords into the chat, leave no record of what they changed, and spend your money in tokens you cannot see. wikiTaTa fixes those four things without asking you to change assistants. You connect it once to Claude, ChatGPT or another assistant that supports MCP connectors. From then on your assistant starts every conversation already knowing your projects, uses your passwords without ever seeing them, writes down what it did, and gets stopped when it is about to do something it should not.
A memory, a vault and a rulebook for your AI assistant. You keep the assistant you already use; wikiTaTa plugs into it.
Under the hood A hosted MCP server (Model Context Protocol — the open standard assistants use to reach outside tools), a Postgres database of “cards”, and a web workspace where you see the same cards your assistant sees.
Four: your assistant forgets between conversations; it handles passwords in plain text; there is no record of what it did; and you cannot see what it is costing you.
Under the hood Session memory loaded at the start of every conversation, a sealed credential vault, an audit trail per action, and per-turn token statistics.
No. Bring the one you have. You keep paying your own AI provider for your assistant.
Under the hood Any MCP-capable client connects over HTTPS with a normal sign-in (some apps limit custom connectors to paid plans). Memory, cards and the server-side safety checks work in every client — for everything the assistant does THROUGH wikiTaTa. What the assistant does with its own built-in tools (its own shell or browser) is outside those checks unless the client supports hooks. Turn-by-turn checks and token tracking need a client that supports hooks (Claude Code today; Codex CLI and Gemini CLI expose the same kind of hooks). The vault works in Claude and Perplexity.
One page about one thing — a project, a decision, a bug, a to-do list, a contact. Your assistant reads cards to get up to speed and writes cards as it works, so the next conversation starts where this one ended.
Under the hood A database row with a title, body, short and long summaries for cheap loading, tags, notes, tasks and links; searchable by meaning, not just keywords.
Built-in memory belongs to one app: it does not carry over to a different assistant. wikiTaTa’s memory is yours, structured, visible, and the same for every assistant you connect.
Those are excellent places for people to write things down and plan together. wikiTaTa is built for the case where an AI is doing part of the work: the assistant reads and writes the workspace itself, under rules, with a record. Many people will use both.
Those turn a description into a working app quickly, and they are good at it. wikiTaTa does not generate apps; it makes the assistant you talk to directly remember, stay in scope and keep your keys safe — whatever you are building, and wherever it is hosted.
Those put an assistant inside your editor. wikiTaTa sits beside them and supplies what an editor does not: memory across projects and months, a credential vault, task tracking, and coordination between several assistants working at once.
It is designed to: an assistant that already knows the project does not need it re-explained, and in clients that support hooks a warning fires when a conversation has grown large and should be restarted, and you can see token use per conversation. We publish our measurement method; we will not quote a percentage until it is measured.
Under the hood Per-turn token capture, context-pressure alerts at set thresholds, tiered card summaries so the assistant loads a line instead of a page.
In our shared database, with each customer’s rows walled off from every other’s.
Under the hood Postgres, with row-level security.
Your assistant cannot: it asks wikiTaTa to use a credential and never receives the value, so a malicious instruction hidden in a web page has nothing to steal. Credentials are encrypted at rest and you enter them through a sealed form, never the chat.
Checks that run around your assistant’s work. In clients that support hooks they run before an action (is it about to print a secret, overwrite someone else’s file, write to the wrong database?) and after (did it verify the thing it says is done?). When a check fails the action is stopped and you are told why.
Under the hood Server-side gates on tool calls, every call logged, plus client-side hooks where the client supports them.
Anyone doing real, ongoing work with an AI assistant — a developer, a freelancer, a small business owner, a team. It is invite-only during early access and free for invited users: request an invitation, set a password, paste one web address into your assistant, and it introduces itself.
We are in the early-adopter stage: invited early adopters pay nothing for baseline services. You pay your own AI provider for your assistant.
Remove the connector from your assistant’s settings — that ends it immediately.
Only what your assistant asks wikiTaTa for during a conversation — the cards and results it requests — which then pass through your assistant like anything else you show it. Secret values are never sent: the assistant receives the result of using a credential, not the credential.
No. wikiTaTa is proprietary software built with open-source components.
wikiTaTa is in private early access. If this is your kind of problem, request an invitation.