Mimir

Mimir is a regulated-grade substrate for AI agents that runs on a user’s VPS and keeps agent activity in a tamper-evident, Ed25519-signed chain. It is presented as a supervisor for agents, with branching and replay tools for recovering a client to a known-good state and for working with decisions after they have been made.

Its core functions are branching, replay, and export. Mimir can fork any audit point at datom granularity, let a different decision run in a parallel timeline, and commit only the branch that wins. It can also re-execute a decision deterministically and intervene at any step, with replay described as byte-identical. For compliance work, it can produce an EU AI Act Article 12 evidence package signed end-to-end with Ed25519. The product also states that every action, read, write, and halt taken by agents is sealed in a tamper-evident chain and retrievable forever.

The system is described as type-checked, AST-gated, and schema-validated. It spawns specialist agents on demand, learns from their work, and supports a tiered model for access control, including PUBLIC, SAFE, SENSITIVE, and PII. The page also says it can branch and replay any decision, coordinate children through software-transactional memory, and audit everything end to end across the fleet. Mimir is built on persistence-os, described as a bitemporal, effect-typed substrate in which facts are immutable datoms, actions are effects, and plans are EDN ASTs.

Mimir is self-hosted on a VPS, and the page says data never leaves the tenant. Its pricing includes a free self-hosted OSS tier under Apache 2.0, Mimir Pro at $29 per month for one operator, Mimir Team at $199 per month for five seats, and an EU AI Act Compliance plan at $2,500 per month per regulated entity. The substrate component is licensed under AGPL-3.