RESEARCH & GOVERNANCE PLATFORM

Deterministic Governance for Multi-Agent AI

AAF (Agents Autonomous Factory) is a research platform for multi-agent software development in which large language models act as typed proposers inside an explicit symbolic control loop, while state, transitions, tests, side effects and completion are owned by a deterministic control layer. The architecture constrains model autonomy with fail-closed control paths and evidence-gated completion workflows.

Read Architecture Contact AAF Labs

Multi-Agent Orchestration & Control Plane

Bounded execution, role capabilities, and a single source of operational truth

AAF orchestrates multiple AI actors (interactive panels, autonomous agents, single-task workers) against a single operational source of truth. AAF's agent-facing control plane requires actors to access the database and infrastructure only through approved tool boundaries (MCP).

Each control-plane call is checked against the caller's role and the command's required capability. This lets the platform combine model-driven proposals with bounded, auditable execution. AAF also maintains a robotics research track (Project ARC) as an active exploratory direction.

Memory Matrix Architecture

AAF uses four distinct memory layers covering project knowledge, temporal continuity, operational state, and analytics.

CMM — Contextual Memory Matrix Tier 1 // Knowledge
Knowledge & Retrieval
What the project knows and how knowledge is connected: project documents, code relations, and semantic graph retrieval (CGRAG).
HTMM — Hierarchical Temporal Memory Matrix Tier 2 // Temporal
Temporal Operational Memory
What happened, when, and how decisions evolved over time, preserving decision history and rejected alternatives across session epochs.
OMM — Operational Memory Matrix Tier 3 // Relational
Transactional Operational SSOT
What is true right now — projects, tasks, ownership, workflow states and runtime identity; the single source of truth for operational state.
ADMM — Analytic Data Memory Matrix Tier 4 // Analytics
Audit Streams & Curated Datasets
A high-volume analytics layer over runtime logs, tool calls, and traces, used to produce curated, versioned datasets and reward signals.

Unified Retrieval Router — evidence from the right memory layer

Route the question. Preserve the evidence boundary.

AAF does not treat every kind of memory as one undifferentiated pool. The Unified Retrieval Router (URR) is the evidence-aware entry point that directs a query to the relevant layers: current operational state in OMM, contextual knowledge in CMM, and temporal history in HTMM. It returns evidence together with its source type and provenance, so the control layer can distinguish what is true now, what the project knows, and what happened over time.

The current routing path prioritises OMM, CMM and HTMM. ADMM contributes curated analytic evidence and datasets for evaluation and reward workflows. The broader learning loop is still being completed, so URR operates as evidence-aware routing.

Deterministic, Evidence-Gated Execution Harness

Symbolic control graphs with a fail-closed safety contract

Allowlisted Actions FAIL-CLOSED
A tool runs only if it is explicitly allowed by the caller’s capabilities; otherwise, the step fails closed.
Typed Parameters TYPED
Action inputs and model proposals are validated against strict type schemas before execution.
Idempotent Fenced Side Effects IDEMPOTENT
Supported mutating ProviderAction paths use stable idempotency keys; ambiguous outcomes are reconciled.
Evidence-Gated Completion VERIFIED
Completion gates evaluate concrete, reproducible test evidence; worker-reported completion remains subject to independent review.

Research Programme & Directions

Investigating robust neurosymbolic agency, long-horizon operational continuity, and formal verification

Controlled LLM Agency ACTIVE RESEARCH
Evidence-gated reasoning scaffolds and fail-closed supervisory control for LLM execution.
Temporal Operational Memory ACTIVE RESEARCH
Chronicle compilation, narrative summarisation, and temporal pruning for long-running multi-agent clusters.
Neurosymbolic Procedural Substrates ACTIVE RESEARCH
Synthesising probabilistic language model proposals with deterministic symbolic control graphs.

Research & Technology Partnerships

For academic collaboration, architecture consultations, or technology partnerships, reach out to our research team.

Email AAF Labs