Synthadoc memory profile

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

Comparison rationale

This profile classifies accepted records, not a fresh source inspection. All positive findings are implementation wiring or the explicitly afforded external instruction-loading path. No source was read, no new record or quotation was added, and no observed-run or causal basis is inferred. Every axis remains partial because the accepted inventory includes unresolved parts; a strong local witness does not settle those alternatives.

Objects, paths and uncertainty

The page body is natural-language knowledge; parsed frontmatter is not uniformly knowledge. Lifecycle fields enforce retrieval eligibility, warnings can veto automatic resolution, and routing/cache/recovery fields select later operations. Text/aliases and vectors also supply ranking inputs. SQLite persistence and vector access both apply to stored embeddings, independently of their distributed-parametric form. These distinctions follow MEM-OBJ-page-metadata, MEM-OBJ-search-vectors, EPI-RTE-warning-gate and MEM-BAP-retained-content. Ordinary model weights and fixed workflow instructions are not accumulated memory in this profile.

Known automatic compilation and scaffold regeneration do not resolve initial authorship or admission. API content suppliers, candidate callers and workflow confirmation clients are not authenticated humans in these records. The separately documented operator-confirmed CLI masking branch supports manual admission; automatic background masking supports automatic admission. A confirmation wrapper does not license a system-wide human-review finding. See MEM-RTE-scaffold-guidance, RT-RTE-candidate-admission, RT-RTE-client-write, WF-RTE-content-health and MEM-RTE-sensitive-mask.

Curation and read-back

Overflow summarization consolidates retained turns even though the raw log remains. That transformation is not proof of summary-fed answering: the inspected handler supplies the recent raw tail, while the stored summary/count govern overflow control and notices. Overview has another known producer without an established semantic answering consumer. No negative trace-learning finding is inferred from either delivery gap. See MEM-RTE-session-context and MEM-OBJ-overview.

General compilation is not automatically synthesis in the controlled sense of creating a claim absent from its inputs. The accepted epistemic account leaves its semantic relation indeterminate. Existing-entry revisions, warning/lifecycle withdrawal, tier promotion and explicit session compression have narrower supported curation mappings. Structural rescanning after a content-health write does not veto the write or prove factual preservation. See RT-RTE-source-ingest, EPI-RTE-citation-annotation, EPI-RTE-lifecycle-admission and WF-RTE-content-health.

Requested wiki/context/export/cached-answer retrieval is pull. Ranking inside that chain is not another unsolicited delivery. Automatic session-history enrichment and standing scope/control supply are separate push operations. The session ID actually selects retained turns before coarse recency/budget filtering; a requested rollback slug alone does not establish identifier push. Likewise BM25/vector relevance does not establish an inferred push signal merely because a query uses an automatic ranker. See RT-RTE-wiki-query, RT-RTE-context-export, MEM-RTE-session-context and MEM-RTE-health-history.

Trace qualification and unions

The imported-transcript chain supports trace learning in the profile's write-route sense: recorded session turns feed automatic durable page production, and later queries can consume those pages. The input experience has learning authority at the update consumer; the resulting page has knowledge authority downstream. Neither classification establishes improved capacity, self-improvement, weight training or verified source claims. The recorded extraction excludes tool/thinking blocks, so session-logs does not imply tool-traces or trajectories. See MEM-RTE-session-ingest, RT-RTE-source-ingest and RT-RTE-wiki-query.

Recovery snapshots, raw evaluation outcomes and scan counters have concrete consumers but are not silently upgraded into additional qualifying trace-learning transformations. Those units retain unresolved qualification/source rather than an unsupported no. Derived system unions therefore retain the wired yes and session-logs witness without claiming complete coverage or treating opaque alternatives as absent. Backup manifests and readable summaries do not resolve native provider checkpoints or arbitrary hook-written payloads. See MEM-RTE-health-history, EPI-RTE-faithfulness-audit, MEM-RTE-sensitive-mask, MEM-RTE-backup-restore, RT-RTE-provider-delegation and RT-RTE-extension-loading.

Canonical citations

Supersessions are resolved to the page-content and faithfulness-cache objects declared in memory; citation-audit routes resolve to the epistemic declaration; cross-wiki and sensitive-mask routes resolve to memory. These affect the object/form/storage/authority and audit/routing/maintenance citations, not their evidence strength. Their complementary findings survive reconciliation. The body/metadata, sidecar/extraction containers, generic/specialized workflow routes and the two rollback mechanisms remain distinct. No supersession supplies a missing trace-learning chain or upgrades an unobserved route. The corresponding canonical references are MEM-OBJ-page-content, MEM-OBJ-faithfulness-cache, EPI-RTE-faithfulness-audit, MEM-RTE-cross-wiki and MEM-RTE-sensitive-mask.