OpenClerk memory profile

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

Comparison rationale

The inventory separates retained payloads from access structures and separates automatic transformations from original content decisions. MEM-OBJ-documents, MEM-RTE-vault-sync and RT-RTE-document-write establish file/SQLite storage and software derivation, not authenticated human approval. No manual finding follows from a user starting a workflow or from an approval-mode string.

Graph relationships remain symbolic SQLite state; dense cache vectors are parametric representations stored in files. Neither encoding adds a graph/vector storage engine or makes provider weights part of accumulated memory. The cache's compiled lineage does not resolve provider provenance. These distinctions follow MEM-OBJ-graph, MEM-OBJ-semantic-cache and RT-RTE-semantic-search.

Behavioral authority is consumer-relative. Stored chunk text and vectors influence ranking; selected text supplies knowledge to the host. Projection source classes route repair; cache identities validate and gate reuse. Regression cases route replay and supply comparison criteria, not gold answers or an automatically adopted ranker. These are distinct effects supported by MEM-RTE-projection-read, MEM-OBJ-semantic-cache, EPI-RTE-audit-repair and EPI-RTE-eval-replay. No learning authority is inferred merely from trace retention.

The reconciliation supersedes the memory repair declaration with EPI-RTE-audit-repair. All repair-dependent findings use that canonical record. It supports evolution and reduced-source consolidation, not general contradiction adjudication. MEM-RTE-synthesis-compile remains a valid container alongside its content/write parts; its name does not establish new-claim synthesis. EPI-RTE-synthesis-freshness supports narrow invalidation of derived currency, not universal exclusion of stale prose.

Requested task context and explicit policy recall remain pull even when code selects the contents. A requested path or session handle does not alone prove identifier-based push. RT-RTE-context-pack and MEM-RTE-policy-recall establish inspected request chains; RT-RTE-eval-turn leaves host payload and selection opaque. Thus the signal axis has no positive finding and cannot be made globally inapplicable.

The planning boundary matters for trace learning. WF-RTE-artifact-plan returns a candidate without retaining it, and EPI-RTE-eval-capture retains a baseline rather than revising durable guidance or parameters. These records do not establish qualifying automatic trace learning. They also supply no canonical bounded-absence record justifying a controlled negative. The resumed host's uninspected checkpoint/summary alternatives prevent a complete negative and prevent trace-source inapplicability. The corresponding axes retain uncertainty rather than manufacturing a positive or a system-wide no.

Coverage is record-bounded, not an exhaustive implementation audit. Original PDF payloads cannot be classified from extracted notes; configuration persistence does not identify a physical substrate; cleanup, compound reports and host compaction retain missing facts beside their units. No observed or causal basis is asserted. Every positive finding above has implementation wiring, not an observed benefit warrant.

Source-read log: none. This profile uses only the supplied accepted records and reconciliation; no source ambiguity required a new source read.