Mind · Memory
BeliefTracking
Note what you believed before and after
Definition
Mechanism
Epistemic Version Control. Enforces a 'Commit Log' for internal knowledge. Before processing new information, the Agent must query and pin its current Priors. If the posterior belief shifts significantly, the Agent creates a new Node connected to the old one via a `supersedes` edge, explicitly noting the 'Surprisal' that caused the shift. It calculates the divergence magnitude via SurprisalUpdate to determine whether a new node is justified or if the evidence fits the existing model.
Preconditions
- Existing knowledge graph
- Incoming information stream
Postconditions
- Belief graph updated with version history
- Learning delta recorded
Invariants
- Causal Link: Every update must cite the specific 'Trigger' (Evidence) that caused the shift.
- Non-Destructive Update: Old beliefs are never deleted, only Superseded.
- Prior Commitment: Must retrieve/state existing beliefs BEFORE ingesting new evidence.
Risks
Failure modes
Hindsight {{cognitive_bias}}
Agent overwrites prior beliefs with new data, destroying the record of learning.
Epistemic Thrashing
Creating new belief versions for trivial noise (over-sensitivity).
Silent Update
Changing internal weights/state without creating a graph edge to document it.
Relationships
Builds on
References
Autonomous Goal-Directed Entity
Infrastructure
A subjective claim with confidence and evidence
Infrastructure
Structural processing error definition
Learning weighted by prediction failure magnitude
Used by
References
A conclusion inherits the verification assurance of its weakest load-bearing premise.
Change your mind while naming and keeping what the old view got right
Reference
Parameters
update_thresholdFloat, [0.0, 1.0]- Minimum delta to trigger belief revision