Why foundry-ai exists

AI coding agents forget. They forget the master context. They forget the constraints you spent an hour explaining yesterday. foundry-ai is the layer that makes them remember — and that audits what they propose before it touches your repo.

Every senior dev I know has had this conversation with an AI agent:

“I told you yesterday to never edit the migration files directly.”

The agent didn’t remember. It couldn’t have. Its memory ends when the session does.

foundry-ai is the layer between you and that forgetting. A small Rust daemon, a SQLCipher-encrypted echo store, and an MCP server that exposes the persisted context back to whichever agent you’re using today.

What it actually does

Captures every prompt, every tool call, every session — in an encrypted local store that survives across agents.
Surfaces a master_context — the project’s purpose, non-negotiables, and stack — that the agent reads on every session start.
Exposes specs as a first-class unit of agent work, with a lifecycle the agent can drive: draft → audited → executing → done.
Audits proposed work through the agent’s own LLM subscription — foundry-ai itself never spends a paid token.

What it isn’t

Not a hosted SaaS. The store is on your machine.
Not a replacement for your AI agent. It wraps Claude Code, Cursor, and any MCP-aware tool.
Not a marketplace. There’s nothing to subscribe to.

If your agent is the brain, foundry-ai is the hippocampus and the prefrontal cortex.