Halofy

Halofy is an open-source access and governance layer for AI agents inside an organization. The project gives agents one governed surface for identity, policy, context, provenance, audit, and signed erasure instead of letting every agent invent its own memory and permission model. The repository describes MCP and HTTP access, Postgres or pgvector-backed operational context, an encrypted Git-versioned cold tier, and a frameworkless control room for managing teams, policies, sources, conflicts, retrieval quality, audit, alerts, skills, and connected apps. The strongest use case is organizational agent memory. A team can connect sources such as files, Postgres, Obsidian, or manual CSV imports, then require reads and writes to pass through server-side identity, namespace ACLs, roles, and policy boundaries. Halofy records provenance and append-only audit history, handles entity resolution and dedupe, and supports signed erasure certificates so removal can be verified. That makes it a fit for teams that want agents to share context but cannot accept unmanaged memory stores or silent data retention. The developer experience is intentionally local-first. The README documents a kernel demo that runs with embedded PGlite, a deterministic stub LLM, and no network access. It also describes an API and control room path for teams that want a managed internal service. Public operations include memory write, read, search, assembly, fault handling, stats, forget, policy, sharing, pinning, export, and manifest actions. Builders can use the driver kit to connect new sources while keeping every write on the same governed path. Halofy is not a simple chatbot template. It is infrastructure for teams that are already serious about agent governance, traceability, and context lifecycle management. Small prototypes may not need the policy and audit surface. Larger teams working with sensitive customer, product, or operational data may value the shared identity model, signed erasure, and offline-ready test lane. The repository is AGPL-3.0-or-later, so the public version is free to run under that license. Organizations should review the license and the self-hosting model before adopting it in a commercial environment. No public hosted pricing was found in the GitHub README during this run. A sensible rollout starts with a narrow namespace, a small set of trusted sources, and a few agents that need the same context. Teams should define roles, read scopes, write scopes, and deletion rules before importing broad company data. Once the kernel, audit trail, and retrieval quality are working for that slice, more sources and agents can be added. That staged approach matches the project's positioning: Halofy is not just memory storage, it is the control boundary around memory. The signed erasure and audit features are especially relevant for organizations that need proof of what an agent saw, changed, or forgot. Provenance helps reviewers understand where a piece of context came from. Append-only history helps incident review. Driver conformance and deterministic tests help teams add connectors without weakening the governance model. Those details make Halofy more complex than a simple RAG stack, but they also make it more appropriate for multi-agent work over sensitive organizational knowledge.