Glossary: memory, persistence, remanence, and correction

Glossary: memory, persistence, remanence, and correction

This family groups the terms that explain what survives after an answer, a correction, a retrieval event, or an agentic action.

The central distinction is simple: a generated response is temporary, but its assumptions, citations, memory objects, retrieved fragments, and external echoes can become durable state. Governance must therefore cover persistence, not only generation.

Canonical terms

• Memory governance governs persisted states and prevents inference fossilization into fact.
• Agentic memory names reusable state that can guide later responses, tool calls, delegations, or actions.
• Memory object defines the typed unit of persisted state that must carry source, authority, scope, freshness, and invalidation conditions.
• Persistent assumptions identify provisional inferences that survive their original context and later guide interpretation or action.
• Controlled forgetting governs invalidation, archiving, deactivation, or non-reuse without silent deletion.
• Stale-state handling governs outdated or potentially outdated state before reuse.
• Surviving authority identifies authority that continues to influence interpretation after it should have lost priority.
• Interpretive remanence names the return of old interpretations after canonical correction.
• Interpretive inertia names the resistance of an established interpretation to correction.
• Version power governs which version of a statement, page, or corpus can impose itself over time.
• State drift identifies divergence between a real current state and the state returned by AI.
• Correction budget estimates the effort required to make a corrected interpretation prevail.
• Resorption describes the neutralization or absorption of obsolete interpretations.
• Correction resorption governs the convergence process through which a correction stops old states from governing outputs.

Operational sequence

1. Identify whether the output created a memory object, assumption, citation, trace, or reusable state.
2. Type the persisted element as fact, preference, assumption, authorization, observation, or historical residue.
3. Attach source, authority, timestamp, freshness perimeter, and invalidation rules.
4. Detect stale-state handling needs before reuse.
5. Apply controlled forgetting when a persisted object should no longer govern outputs.
6. Track remanence, inertia, version power, and surviving authority after correction.
7. Measure correction resorption until the old state no longer activates by default.

This sequence prevents a common failure of stateful systems: treating memory as reliable simply because it is available.

Related pages

• Memory governance
• Glossary: drifts and interpretive inertia
• Glossary: agentic execution and transactional control
• Glossary: RAG, retrieval, and documentary chain

Phase 10 routing layer: inference, arbitration, indeterminacy and fidelity

This page now routes inference-control questions toward the phase 10 canonical layer: interpretive error space, free inference, default inference, arbitration, indeterminacy, and interpretive fidelity.

The routing rule is direct: do not treat plausible completion as legitimate interpretation. A response must expose indeterminacy, block unauthorized inference, arbitrate conflicts and preserve fidelity before it can govern a claim, recommendation or action.

Phase 12 routing layer: debt, maintenance, and deprecation

This page now routes maintenance and long-term correction questions toward the phase 12 canonical layer: semantic debt, canon maintenance, interpretive maintenance, maintenance burden, correction backlog, deprecation discipline, canonical refresh cycle, and obsolescence control.

The routing rule is direct: a canonical corpus does not remain reliable through publication alone. It requires maintenance, status control, deprecation, backlog management, artifact synchronization and correction resorption.

How to read this lexical family

This family describes the afterlife of meaning. Once a claim has appeared in an output, a memory, an index, a citation, a summary or a retrieved fragment, it may continue to influence interpretation after it should have been corrected. That survival is not the same as current authority.

The terms form a correction discipline. Memory governance asks what should persist. Agentic memory and memory objects describe what can be stored. Persistent assumptions explain what continues to shape outputs. Controlled forgetting and stale-state handling define how obsolete material should be limited. Correction resorption tests whether correction has actually taken effect.

Typical misreadings

The most dangerous mistake is to treat persistence as validation. A statement repeated by multiple systems may simply be a surviving error. It may have propagated through citations, cached summaries, model memory or user prompts without becoming more legitimate.

Another mistake is to treat deletion as forgetting. Removing a page or changing a statement may not remove the interpretation from external systems. The corrective task is therefore not only publication, but resorption across the surfaces where the old state survives.

Use in audit and routing

Use this family when a corrected claim keeps reappearing, when an agent relies on an obsolete assumption, when a model repeats an old framing or when an external graph preserves a prior state. The audit should identify where the old state survives and what evidence shows whether it is losing force.

For routing, this family supports memory governance, stale-state handling, correction budget, version power and interpretive remanence pages. It is the persistence layer of interpretive governance.