Buddhi AI

Why it exists

AI agents now write code, query databases and summarise Slack, but most companies have no idea what those agents can reach. One contractor’s agent reads the whole codebase; an intern’s bot pulls exec-level financials.

What it does

• Sits between every agent and every memory store, and runs five checks on each request: identity, role policy, scoped retrieval, redaction and an audit log.
• Every answer carries its source, author and timestamp, so numbers come from the record instead of the model.
• When two sources disagree, the one with more authority wins (a decision doc outranks a stale thread) and the dissent stays linked.
• One MCP endpoint for Claude Code, Cursor, Codex and ChatGPT, plus a CLI and a REST API, all under the same policy.
• Read-only connectors for GitHub, Slack, Google Workspace and Notion, and workflows such as cited daily standups and compliance digests.

What it won’t do

• Deny is the default: if a scope isn’t explicitly allowed, the agent gets nothing, and out-of-scope questions are refused, never approximated.
• No anonymous queries: every request carries a specific agent acting for a specific person.
• API keys, passwords and personal data are redacted before anything reaches the agent.
• Connectors only read: it never sends, edits or deletes anything in the tools it connects to.

How it works

PostgreSQL is the authoritative store for memories, policies, users and audit logs, and Qdrant runs vector search filtered by permission. A memory orchestrator sits in the middle, and no code path bypasses the policy engine.

By the numbers