Your docs disagree with each other. The spec says one limit, the runbook says another, and your agents' notes still cite last month's decision. Constrate finds every contradiction across your doc suites, opens the fix in the original document, and hands Claude, ChatGPT, Codex and Cursor only the current, cited version.
Reads Google Docs & Sheets · Word & Excel / SharePoint · Quip · Notion · Confluence · CLAUDE.md / AGENTS.md / agent memory — serves any assistant over MCP or API
Ask "who owns on-call for payments?" and a standard retrieval pipeline hands the model twice the text it needs, including lines that name the old owner. Constrate hands it the answer, with the source and date on every line.
Example from Constrate's synthetic evaluation company; names are fictional.
Newest-file-wins breaks the moment someone keeps discussing a topic after a decision. Constrate works one level down: on the facts themselves.
Read-only sync from every suite, with each document's permissions carried along. Runs in your cloud if you need it to.
Every document becomes claims: subject, value, as-of date, author, source, and how authoritative that source is.
Conflicting claims are reconciled into one current value per fact, with the losing versions kept as history and their documents flagged.
Stale docs get a suggested edit, comment or pull request that a human approves. Your assistants call one MCP tool for compact, cited, permission-filtered context in the meantime.
fewer tokens handed to the model than standard retrieval
of contexts clean and current, vs 23% for standard retrieval
stale lines passed to the model, across 60 questions
Measured on a synthetic 551-document company across five sources, built with known ground truth so every context can be scored without an LLM judge. Real-workspace evaluations with design partners are under way.
We'll run a free stale-docs report on one folder or workspace you choose, and show you what your AI assistants have been reading. We're onboarding a handful of design partners this winter.
Request early access