Context Persistence

Context persistence solves the reset between work sessions. An agent may gather evidence, wait for a person to answer a question, pause while an external system responds, and continue after a source changes. Saving only the last message forces the next run to reconstruct why the work exists and which constraints still apply. A persistent checkpoint records the objective, current state, selected source versions, accepted findings, open questions, allowed tools, budget or time limits, pending external actions, and responsible roles. The next participant can resume from a deliberate operating state instead of guessing from conversation history.

A checkpoint is not copied blindly into the next run. Resumption validates identity, permission, source freshness, approval expiry, pending receipts, and changes to the target object. If a customer, environment, amount, policy, or destination differs, earlier authority may no longer apply. If an external action timed out, the workflow reconciles whether it occurred before retrying. Persistence preserves what was known and authorized at an earlier time; validation determines what remains safe and relevant now. The correct continuation can be refresh, narrow, escalate, or stop.

Persistent context has several lifetimes. Temporary working material may expire when one model call or run ends. Workflow state remains until the case reaches a terminal disposition. Approved company records can persist longer under source ownership and retention rules. Historical evidence may be retained separately for reconstruction. Keeping these lifetimes distinct prevents a draft, private chain of messages, old approval, or temporary tool result from becoming permanent guidance. It also lets the company delete unnecessary material without destroying the compact evidence needed to understand a decision.

Long-running work is normal in companies. Reviews take time, providers respond asynchronously, incidents cross shifts, and a case may reopen after new evidence arrives. Without persistence, people repeatedly explain the state or allow a model to infer it. Important constraints disappear, abandoned branches return, and already completed actions can be repeated. A durable checkpoint reduces this handoff loss. It also makes a different agent or human able to continue the case, lowering dependence on one session, model, or original operator.

Persistence is also a control against invisible authority expansion. A saved approval can look like permission to keep acting indefinitely unless its scope and expiry remain attached. A successful earlier run can appear to justify a different customer or environment. By recording the exact action, subject, amount, destination, version, and time, the workflow can detect a mismatch before continuing. The same mechanism preserves refusals and unresolved conditions, so a new worker does not route around a boundary simply because it was not present in the latest message.

For recovery, context persistence connects technical state with business uncertainty. A queue checkpoint that says completed is insufficient if the target system has not confirmed the change. A durable case record keeps the attempt, correlation identifier, receipt posture, owner, and next verification step together. Operators can resume after an outage without scanning unrelated logs or repeating side effects. This continuity supports reliable service, but it does not guarantee that every dependency is available or that retained context is correct; those conditions remain subjects of validation.

The checkpoint stores a stable workflow and object identifier, objective, current state, source references and versions, accepted findings, unresolved conflicts, prior decisions, authority scope, pending actions, budget, deadlines, owner, and permitted next transitions. It references protected evidence instead of copying secrets or full sensitive payloads. Material fields are typed so a proposal, approval, attempt, provider receipt, verified change, and observed outcome cannot collapse into one completed label. The contract remains understandable outside the original agent transcript.

Before work resumes, the system checks the current identity, entitlement, source freshness, policy version, approval validity, target-system state, and pending external receipts. It compares the checkpoint with events that occurred while the workflow was paused. Conflicts are retained and routed to an owner rather than resolved by the newest timestamp or most fluent summary. Duplicate callbacks and stale workers cannot advance the state. A different agent or person can continue only through the same validation boundary.

Working context, resumable state, durable company records, and historical evidence receive separate expiration and access rules. Customer, employee, financial, security, and contractual material remains purpose-bound when summarized, embedded, cached, or checkpointed. Derived state carries links to correction and deletion requirements. The workflow limits context to what the next role needs and preserves the reason for retention. Long-lived does not mean universally visible, and a technical ability to store context does not establish a lawful or useful retention period.

External attempts retain correlation and idempotency data, responses, errors, and reconciliation posture so resumption cannot casually repeat a consequential action. Corrections create a new version and identify affected downstream context without rewriting history. Exportable state and stable interfaces allow a human operator or substitute worker to reconstruct the case if the original runtime is unavailable. A manual fallback can finish essential work while preserving later reconciliation, and closure settles pending actions, temporary credentials, and retained evidence.

Keeping every prior message available can increase distraction, privacy exposure, cost, and prompt-injection persistence. Old context can remain semantically relevant after losing authority. Persistence should preserve selected state and source references with explicit lifetimes, not replay an unlimited transcript into every run. Summaries help compression but remain derived records that require lineage and correction. The goal is a sufficient checkpoint for the next decision, not maximum historical volume.

An approval, credential, or successful action from an earlier session does not authorize every continuation. Authority is bound to the original scope, subject, action, amount, destination, environment, and time. Resumption may require a current decision when those dimensions change or when policy expires. Persistence protects the evidence of earlier authority and makes its limit visible. It must not become an invisible mechanism for escalating from research to mutation, draft to send, or preparation to release.

A durable checkpoint can faithfully preserve an incorrect source, weak interpretation, or unresolved conflict. Model output can also carry malicious or accidental instructions into later runs. Source classification, instruction-data separation, permissions, freshness checks, and human review reduce risk but do not guarantee correctness or security. High-impact work may require direct validation against the owning system and specialist judgment. Persistence makes continuity inspectable; it does not make every retained statement authoritative.

Consider a hypothetical service incident that begins near the end of an operations shift. An agent helps assemble approved telemetry, recent change records, affected service boundaries, and the current response procedure. The incident owner authorizes one reversible diagnostic action and asks the team to wait for a provider status update. The checkpoint records the observed symptoms, source timestamps, action receipt, unresolved provider state, prohibited customer and production actions, next review time, and the oncoming owner's identity.

Before the next shift resumes work, the workflow refreshes service status, verifies that the diagnostic action did not alter protected state, checks whether the provider response arrived, and confirms that the new owner has the required access. One source now conflicts with the saved summary, so the context assembler preserves both readings and marks the earlier interpretation as challenged. The agent may prepare options, but it cannot send customer communication, make a public status claim, spend beyond the incident limit, or deploy a change without the current authority for those actions.

The provider eventually confirms a condition that supports a different diagnostic path. The owner records the decision, and the checkpoint advances with the new evidence while retaining the earlier basis. If the original agent runtime becomes unavailable, another worker or a human can reconstruct the case from the same compact record and protected source references. The example does not claim faster recovery or service reliability. It shows how persistent context supports handoff, revalidation, and non-duplicative action under changing evidence.

Pause representative workflows before approval, after an external attempt, during a source update, and following denied permission. Resume with the same agent, a substitute worker, and a human operator. Change a source, expire an approval, revoke access, duplicate a callback, and alter the target record while paused. Verify that each participant receives the appropriate state and that continuation refreshes, refuses, reconciles, or escalates instead of blindly executing the saved next step.

Ask an independent reviewer to reconstruct the objective, evidence versions, decisions, authority scope, pending external state, owner, and next valid transition without reading the full transcript. Then inspect whether the checkpoint copies secrets, excessive personal data, or unrelated source content. Test retention, correction, deletion, and supersession across derived summaries and caches. The record should be compact enough to govern and support yet complete enough to prevent repeated or unauthorized action.

Track stale resumes, duplicated effects, abandoned cases, reconciliation time, successful handoffs, correction effort, review burden, storage and model cost, and manual fallback quality. Replace a model or worker and confirm that company state remains coherent. Compare with the prior handoff process over similar cases. A high technical resume rate is not useful if resumed work carries inappropriate context or hides unresolved destination state. Expansion requires both operational usefulness and proportionate lifecycle cost.