Core Concepts

Understanding the building blocks of AMFS.

AMFS treats each agent’s knowledge as a Git repository. Every write is a versioned commit. Every operation is logged on a timeline. Branches isolate experiments. Pull requests let you review changes before merging. You can roll back to any point.

On top of this Git-like foundation, AMFS adds memory intelligence: memory entries store knowledge with a memory type (fact, belief, or experience), copy-on-write preserves full history, provenance tracks authorship, confidence scores evolve based on real-world outcomes, a knowledge graph auto-materializes relationships, and hybrid search blends full-text, semantic, and confidence signals.


At a Glance

Agent writes →  MemoryEntry created
                  ├── entity_path: "checkout-service"
                  ├── key: "retry-pattern"
                  ├── value: { ... }
                  ├── memory_type: fact   ← fact / belief / experience
                  ├── version: 3          ← CoW versioning
                  ├── confidence: 0.85    ← evolves with outcomes
                  ├── branch: "main"      ← git-like branch
                  └── provenance:
                      ├── agent_id: "review-agent"
                      ├── session_id: "abc-123"
                      └── written_at: 2025-06-15T10:30:00Z

              →  Event logged on agent timeline
                  ├── event_type: "write"
                  ├── branch: "main"
                  ├── summary: "Wrote checkout-service/retry-pattern v3"
                  └── details: { entity_path, key, version, confidence }

Read the pages in this section to understand each concept in depth.


Table of contents