BeliefGraph

BeliefGraph is an open framework for persistent reasoning in AI agents.

Quick Start

Clone the repository and run the belief propagation demo.

git clone https://github.com/joshuamlamerton/BeliefGraph cd BeliefGraph python examples/demo.py

Most AI agents today are stateless. Each interaction starts fresh or relies on vector retrieval. This makes it difficult for agents to maintain long-term beliefs, track assumptions, or update conclusions as new evidence appears.

BeliefGraph introduces a structured layer that allows agents to maintain explicit belief networks.

These networks represent:

beliefs
evidence
confidence levels
relationships between claims

Agents update these graphs over time as new information arrives.

---

Core Idea

Instead of storing knowledge only in embeddings or prompts, agents maintain a structured belief graph.

Each belief contains:

• a statement
• a confidence score
• supporting evidence
• relationships to other beliefs

---

Architecture

flowchart TB

A[Agent]

B[BeliefGraph Layer]

C[Belief Nodes]

D[Evidence Links]

E[Confidence Update Engine]

F[External Tools<br>APIs / Data Sources]

A --> B
B --> C
B --> D
B --> E
A --> F
F --> B

Loading

---

Example Belief Node

{
  "belief": "interest_rates_increase_affects_mortgage_affordability",
  "confidence": 0.82,
  "evidence_sources": ["market_data", "historical_trends"]
}

Running the Prototype

Clone the repository and run the demo:

python examples/demo.py

The script demonstrates:

• creating a belief
• adding evidence
• updating confidence using a Bayesian rule

Belief Propagation

Beliefs in the graph can influence other beliefs through weighted relationships.

Example:

interest_rates_increase → mortgage_affordability_declines

When confidence in the first belief increases, the second belief updates automatically.

This is implemented through the propagation engine.

Run the demo:

python examples/demo.py

Contradiction Relationships

Beliefs can also contradict other beliefs.

Example relationships:

interest_rates_increase → mortgage_affordability_declines (support)

interest_rates_increase → housing_demand_strong (contradict)

Support edges increase confidence in the target belief.
Contradiction edges decrease confidence in the target belief.

Run the demo:

python examples/demo.py

---

Roadmap

Phase 1
Belief node schema

Phase 2
Graph update engine

Phase 3
Agent reasoning queries

Phase 4
Multi-agent belief sharing