Mechanism
Governed forgetting is a governed process by which specific lineage entries are deprioritized, not deleted, through a process that is itself recorded in the lineage. The distinction is structural and load-bearing: the lineage entry remains immutable and complete within the lineage record. What changes is the weight the entry receives when the agent's cognitive subsystems evaluate the agent's behavioral history. Governed forgetting is not data loss. It is relevance decay.
Concretely, the weight assigned to a lineage entry in integrity computation, in self-esteem evaluation, and in trajectory analysis decreases over time according to a policy-defined decay function. As that weight falls, the entry contributes progressively less to the agent's active cognitive computations, while remaining fully present and reconstructable in the lineage. The agent prioritizes recent and relevant behavioral history in its ongoing reasoning without ever losing the ability to reconstruct the complete lineage for audit purposes.
Why Forgetting Is Not Deletion
The mechanism modulates computational weight, not the existence of the record. Where a deletion operation would remove a datum and leave the absence of the datum as its only trace, governed forgetting leaves the lineage entry intact and modulates only the influence the entry has when the integrity engine, the self-esteem computation, or the trajectory analysis subsystems read the agent's history. The entry that has been forgotten in this sense is still there to be audited. It simply no longer dominates the agent's present cognition.
This is what allows the agent to move past a stale or superseded portion of its history without erasing it. A deviation that has been corrected, a behavioral pattern that has been overtaken by subsequent events, a portion of history that policy has narrowed out of current scope: each can be down-weighted so that it no longer drives present reasoning, while the complete account remains available to any party reconstructing the lineage.
The Decay Function
The rate and shape of relevance decay are set by a policy-defined decay function. The decay function specifies the rate at which relevance diminishes, the minimum residual weight below which a lineage entry no longer contributes to active cognitive computations, and the conditions under which decay is accelerated or suspended. The function is not fixed by the mechanism: it may be linear, exponential, a step-function, or a policy-custom form, with the applied form recorded as part of the forgetting event.
The minimum residual weight is significant. The decay does not drive an entry's weight to zero and thereby out of computation entirely; it drives it down to a policy-defined floor below which the entry no longer participates in active cognitive computations. The entry persists in the lineage at full fidelity regardless of how far its computational weight has decayed.
Forgetting as a First-Class Governance Event
The decision to deprioritize a lineage entry is not a silent background process. It is a first-class governance event, evaluated by the composite admissibility evaluator before the deprioritization is applied. The composite admissibility evaluator is the combined mechanism through which the confidence governor, the integrity engine, the capability envelope, and other cognitive domain field evaluators integrate their assessments into a single admissibility determination. A forgetting event must clear that evaluation in the same way any proposed mutation must.
The governance evaluation records in lineage the specific lineage entry being deprioritized, the reason for the deprioritization, the decay function applied, the policy authority that authorized the deprioritization, and the conditions under which the deprioritization is reversible. The enumerated reasons include staleness, supersession by subsequent corrective action, policy-directed scope narrowing, and relevance expiration. Because the forgetting event is itself recorded in lineage, the act of forgetting is as auditable as the entry it acts upon.
Relevance Restoration
Because nothing is deleted, governed forgetting is reversible. The mechanism supports relevance restoration: when a deprioritized lineage entry becomes relevant again, its computational weight is restored through a governed reversal event that is itself recorded in lineage with full provenance. An example is a new mutation that raises questions about a behavioral pattern documented by an entry that had been deprioritized. The entry's prior decay does not prevent it from being brought back into active computation, because the entry was never removed, only down-weighted.
Restoration completes the symmetry of the mechanism. Forgetting and restoration are both governed events, both evaluated and both recorded, so the agent's history of what it chose to emphasize and de-emphasize is itself a part of the lineage that can be reconstructed and audited.
The Subsystems That Read the Weight
Governed forgetting takes effect through the subsystems that consult the weighted lineage. Integrity computation, self-esteem evaluation, and trajectory analysis each read the agent's behavioral history when forming their assessments, and each reads it through the lens of the current relevance weights. A deprioritized entry exerts less pull on the integrity field, contributes less to self-esteem evaluation, and counts for less in the trajectory analysis that classifies the agent's overall behavioral arc.
This is the locus of the mechanism's effect. The weight is not an abstract bookkeeping value: it determines how strongly a given moment of the agent's recorded history shapes the agent's present integrity assessment and its corrective behavior. By modulating that weight under governance, the mechanism lets the agent's active cognition track recent and relevant history while the deviation log and the full lineage remain complete for forensic reconstruction.
Distinction from Deletion-Based Approaches
Approaches that satisfy forgetting requirements by deleting records trade away auditability to obtain it. Once a record is gone, the system cannot reconstruct the behavioral history that produced its present state, and a forgetting claim is evidenced only by an absence. Governed forgetting takes the opposite structural position: the record is preserved in full, the forgetting acts on computational weight rather than on existence, and both the forgetting and any later restoration are recorded as governed events with full provenance.
The consequence is that the agent can deprioritize history for the purposes of its own ongoing cognition while remaining fully reconstructable for audit. The agent does not have to choose between moving past stale history and preserving an auditable record of it. The mechanism is built so that the agent keeps both.
Disclosure Scope
Governed forgetting, comprising the deprioritization rather than deletion of specific lineage entries through a policy-defined decay function, the decay function's specification of decay rate, minimum residual weight, and the conditions for accelerating or suspending decay, the evaluation of each forgetting event by the composite admissibility evaluator as a first-class governance event, the recording in lineage of the deprioritized entry, the reason for deprioritization, the decay function applied, the authorizing policy authority, and the reversibility conditions, and the governed relevance restoration of a previously deprioritized entry, is disclosed in the cognition filing (U.S. Application No. 19/647,395 and its international counterpart). This article describes that disclosed mechanism. The scope extends to the decay function forms enumerated, namely linear, exponential, step-function, and policy-custom, and to embodiments in which integrity computation, self-esteem evaluation, and trajectory analysis consume the weighted lineage, provided the lineage entry itself remains immutable and complete and the forgetting acts only on the computational weight the entry receives.
