Stash memory profile
Type: agentic-system-analyses/types/agentic-system-memory-profile.md
Comparison rationale
The comparison is architectural, not an operational success report. The accepted inventory separates the retained objects from provider weights, compiled instructions and release machinery. PostgreSQL is the substrate; pgvector is a retained vector surface. Relational edges do not independently establish graph storage. Dense vectors receive the parametric form, not natural-language form, even when generated from readable text. See MEM-OBJ-memory-vectors, MEM-OBJ-embedding-cache and RT-CMP-openai-embedder.
Canonical citations use the memory object declarations and pattern/failure routes selected by reconciliation. Superseded epistemic object declarations, pattern extraction and failure review are not cited as independent witnesses. Their identity amendments change citation destinations, not evidence strength. The combined edge, hypothesis, goal and withdrawal routes remain valid overlaps, not aliases of their narrower epistemic routes or of the runtime consolidation container.
Curation and epistemic warrant answer different questions. Summarizing retained episodes is consolidation even when actual semantic faithfulness is unestablished. Hypothesis confirmation promotes content without necessarily accepting it against evidence; the automatic scan and direct caller confirmation have different acceptance paths. Neither extraction wording nor a command named synthesis establishes a new claim absent from its inputs. This prevents a positive synthesize classification, not the positively wired consolidation, promotion and state edits. See EPI-RTE-fact-synthesis, EPI-RTE-hypothesis-assessment and EPI-RTE-hypothesis-control.
Read-back is split by operation. External recall and typed reload are requested pulls. Maintenance selects and supplies batches to its model without that model requesting those retained rows. Those pushes have coarse windows, embedding-based episode grouping and exact entity/property contradiction matching. Background persistence alone is not a foreground context push. No finding treats query-vector pull retrieval as inferred-embedding push. See RT-RTE-recall-read, MEM-BAP-maintenance-input and EPI-RTE-contradiction-review.
Experience reports can feed qualifying automatic durable writes without authenticated logs or measured learning benefit. Failure-review outputs and episode summaries have later consumers. Their original inputs, however, are not established as conversational turns, tool execution records, discrete environmental events or linked trajectories. Thus local trace-learning positives do not fill trace_source by inference from the word episode. Generic caller content also prevents an exhaustive trace-learning conclusion. See MEM-RTE-failure-derivation, MEM-OBJ-episode and RT-RTE-consolidation-run.
The unresolved theory-builder condition-3 label disagreement has no effect on these memory values. Neither label supplies a criticism instance or a synthesis/benefit warrant. State-dependent scans and retained rejection reasons remain supported independently. See MEM-RTE-hypothesis-maintenance and EPI-RTE-hypothesis-assessment.
Coverage remains partial where original admission control, upstream provenance, pattern/getter consumers and additional settings uses remain unresolved. There are no invented bounded-absence records and no global trace-learning negative. Known schema inventory does not imply a complete audit of every constraint or consumer. Stored validity is especially not equivalent to enforced retirement: ordinary recall still admits ended facts. See MEM-OBJ-settings, MEM-RTE-typed-read and MEM-RTE-fact-withdrawal.
Source-read log: no target source reads. This profile classifies only the supplied accepted records and their annotations; it adds no records, quotations, execution evidence or independent source findings.