An open-source project makes agents prove they erased data

Halofy went public August 22 with a governance layer for AI agents that can prove when it deletes their memory.

The AGPL-3.0, TypeScript project sits between MCP servers and an agent's memory, adding access control, role-based permissions, tenant isolation, data provenance and audit logs. Its most unusual feature is signed erasure, a verifiable record that data an agent held was destroyed for good. Compliance teams ask for that kind of proof, and most agent stacks have no way to give it.

The halofyai/halofy repo lists pgvector and model context protocol among its topics. That places it as a layer watching what agents remember, not a runtime hosting them. It passed 300 stars in the four days since release.

My bet is that erasure without a signature won't hold up in an audit. Running agents against regulated data will soon require proof that deleted context stayed deleted, and almost nothing in the agent ecosystem can produce that proof yet.