schema_version: infospace-bench.memory-context-evaluation.v1 evaluated_at: "2026-05-15T00:20:00Z" workplan: IB-WP-0017 profile_id: infospace-agentic-memory-pilot graph_id: infospace-agentic-memory-graph selection: restart-context-selection.yaml questions: - id: restart-quality prompt: Can a later agent resume Wealth/VSM review with the selected package? score: 4.2 max_score: 5.0 result: pass - id: provenance-review prompt: Can a reviewer see why each memory item exists? score: 5.0 max_score: 5.0 result: pass - id: budget-realism prompt: Does the selection fit a compact restart budget? score: 4.0 max_score: 5.0 result: pass - id: noise-control prompt: Does the package omit low-value trace detail? score: 3.8 max_score: 5.0 result: watch metrics: restart_quality_score: 4.2 provenance_coverage_ratio: 1.0 selected_node_count: 7 expected_item_count: 8 context_package_budget_max_tokens: 1200 live_llm_required: false findings: - id: finding.restart-risk summary: Decision-only memory is not enough; restart packages need the active source/entity neighborhood. - id: finding.neighborhood-improves-review summary: The Division of Labour source and entity references make the boundary decision actionable for review. - id: finding.profile-gap summary: Profile retention intent is useful, but acceptance thresholds remain application-level metrics in infospace-bench. recommended_contract_changes: markitect-tool: - Add an option for timestamp-stable context package fixture output to simplify cross-repo golden files. - Document when selected events should become package items versus metadata. kontextual-engine: - Import Markitect graph/profile envelopes and persist runtime audit events separately from contract events. - Keep durable write plans review-gated and export Markitect-compatible package input envelopes. infospace-bench: - Keep memory quality metrics and pilot corpora here. - Do not store user memory, credentials, or runtime graph state in an infospace.