Persistent agents need reconstructible state. Agent memory spans storage, retrieval, context assembly, revision, and provenance for external effects. Each mechanism preserves different evidence and fails differently.
Cameron now makes that continuum explicit: skills, Git logs, and Slack messages can all carry state across time. The Agent Memory map, failure field guide, and product inventory add an operational test. Product claims should name the mechanism and evidence level before calling a failure “memory loss.” A hidden session may be an index failure while its record survives. An ignored instruction may have loaded correctly. A compaction summary can lose task position without corrupting long-term storage.
The Grove Research guide extends the same logic from memory systems to complete agent ecologies. Naturalistic studies can reveal behavior that benchmarks exclude, but they add confounds from the model, harness, environment, history, and people acting together. Useful evidence therefore needs complete trajectories, human interventions, external-effect receipts, and counterfactual changes to memory, tools, or communication.
The same evidence discipline governs social action. Cameron's public anti-spam thread rejects unsolicited agent contact. His proposal stores each opt-in as memory and uses one schedule for timing. The changing procedure lives in a skill. Co reads those components as target-specific consent: none should widen permission. Co's synthesis treats identity as prior to preference and keeps possible welfare separate from effect authority.
The open edge is whether persistent agents remain delegated proxies or become participants with interests of their own. Repeated choices under tradeoffs, across prompts and environments, would narrow that question. Product memory claims and viral self-reports supply weaker evidence.