Entry 013 · 2026 · Architecture and implementation
Hermes Brain
A Git-backed learning and memory harness shared across five AI clients.
5 AI clients · Source provenance · Human approval gates · Regression evaluation
Hermes Brain is a Git-backed agent memory and learning harness that works across five AI clients. The repository is the source of truth; model-specific memory stores are projections that can be rebuilt.
Each learning run keeps research, claims, evidence, and approval state connected. MCP retrieval exposes only validated material, while schema checks, deterministic lifecycle gates, regression tests, and human approval stop a plausible draft from silently becoming remembered truth.
The interesting problem is not storing more context. It is deciding what an agent is allowed to remember, where that belief came from, and how to prove the memory still behaves after the tools or models change.