beads_rust memory profile

Type: agentic-system-analyses/types/agentic-system-memory-profile.md

Comparison rationale

The inventory follows the accepted memory boundary, not the executable's entire storage or permission surface. The same issue has prose for interpretation and symbolic state for fixed readiness and transition consumers. SQLite and files therefore coexist without treating JSONL encoding as a separate representational form. Local events and scoped gate evidence are not ordinary export payloads; recovery copies do not reproduce every intermediate change (MEM-OBJ-issue, MEM-OBJ-event, MEM-OBJ-history-backup, EPI-OBJ-gate-result).

The reconciliation supersedes the memory gate object and reporting route with EPI-OBJ-gate-result and EPI-RTE-gate-report. All profile gate citations use these canonical destinations and the separately declared checking/admission routes. This changes citation identity, not the warrant: supplied verdicts govern report sufficiency, not independently verified task truth. Effective-history selection supports invalidation, while the guarded consumer supports validation and enforcement (EPI-RTE-gate-check, EPI-RTE-gate-admission).

Read-back is split by operation. The requested child/detail or prompt return is pull. The optional ancestry selector additionally resolves parent/root identities and supplies their governing text; this local automatic enrichment is push even though an upstream child request triggers it. It is not an independent scheduler, host-eviction detector or evidence of instruction compliance. Identifier classification applies to the actual matched ancestor blocks, not an ID printed in a catalog. Structured update and MCP detail do not inherit the text-update enrichment guarantee (MEM-RTE-inherited-context, MEM-RTE-state-read).

Curation preserves its evidence layers. Existing-entry updates and configured snapshot deduplication/forgetting are wired; backlog-polish editing is afforded to an external agent. Neither transient projected prompt context nor physical compaction establishes retained semantic consolidation. Automatic event recording does not by itself establish trace learning, and explicit caller writes do not establish manual agency. Those distinctions leave unresolved axes unresolved rather than manufacturing negative findings (MEM-RTE-guided-context, MEM-OBJ-history-backup, MEM-CLM-compaction-scope, MEM-RTE-audit-record).

Complete classification is prevented by unaudited auxiliary evidence/provenance, recovery branches, instruction-distribution alternatives and the audit log's later semantic consumer. The named unknown units retain these limitations; strong local witnesses do not upgrade them. Installed project guidance has an afforded future instruction consumer, not a demonstrated host delivery policy. No finding is observed or causally supported, because the accepted reports supply implementation and doctrine rather than execution or comparative outcomes (RT-RTE-agents-update, RT-BAP-agent-instructions, MEM-RTE-state-withdrawal, MEM-RTE-audit-record).

Source-read log: none. Classification used only the supplied accepted members and reconciliation; no additional source evidence or records were introduced.