Belief Context Graph
A confidence-aware context graph for long-horizon AI agents that tracks not just what was said or retrieved but what an agent should believe, at what confidence, from which evidence, and whether it should act on it.
Belief Context Graph is a memory substrate that goes beyond retrieval-style agent memory, conversation history, vector search, GraphRAG, or trace logs, by tracking belief questions: whether a fact is still valid, how reliable its source is, whether it conflicts with other evidence, and whether a past belief was later proven wrong by an outcome.
It segments an agent's trajectory into extracted, structured beliefs and links them into a connected graph, computing a deterministic, auditable posterior confidence for each belief from its initial confidence, evidence confidence, and confidence derived from related beliefs, so relation weights and activation thresholds propagate support or contradiction in a reproducible way rather than through another opaque model call. Every belief keeps an exact-offset reference back to the conversation turn it came from, and beliefs carry timestamps recording when they were formed and how they evolved across a session.
It requires Python 3.11 through 3.13, and its optional reference terminal agent adds a runnable example of the graph in use.