Secure LLM-Wiki memory profile

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

Comparison rationale

This profile classifies accepted implementation records, not a deployed wiki instance. Wiring establishes delivery and code effects; no finding asserts observed model compliance, factual accuracy, attack resistance or benefit. The retained inventory is separated into text, control metadata, numerical vectors, diagnostics, editable policy and history because their consumers differ (MEM-OBJ-claim-text, MEM-OBJ-claim-metadata, MEM-OBJ-claim-embeddings, MEM-OBJ-gate-records, MEM-OBJ-trust-rules, MEM-OBJ-wiki-history).

Reconciliation supersedes the epistemic claim-text and embedding objects with the memory objects cited here. It does not supersede the broader epistemic metadata container by its finer memory parts. No value relies on a superseded citation. Indeterminate extraction fidelity is not evidence of synthesis: the extraction route establishes source-derived compilation but no concrete novel claim or transformation over retained source sessions (RT-RTE-extraction-parsing).

Vector coordinates count as distributed-parametric form and a vector substrate even though their serialization is a JSON file. Their actual authority is at admission: maximum cosine selection ranks matches, thresholds supply validation criteria, and the gate enforces rejection/escalation. This is not question-conditioned retrieval. Query text instead supplies knowledge; its location in a system string does not establish instruction authority (MEM-OBJ-claim-embeddings, RT-BAP-query-evidence).

Operator initiation is not manual admission of each extracted claim. Manual edits and confirmed destructive maintenance are separate mechanisms; generic storage caller identity cannot determine whether a human approved content. The same distinction prevents arbitrary caller replacements from resolving lineage or automatic trace learning (MEM-OBJ-claim-text, RT-RTE-operator-maintenance).

The explicit query-feedback absence warrants only a local trace-learning negative. Source-derived context is retained, but records do not establish whether original ingested documents contain qualifying activity traces; retained review outcomes and Git messages are not automatically transformed into new guidance. Opaque provider updates remain unresolved. Therefore the partial learning inventory has no system-wide negative, and its source axis is not globally inapplicable (MEM-ABS-query-feedback, MEM-OBJ-gate-records, MEM-OBJ-wiki-history, RT-CMP-review-model).

Requested startup loading remains pull even when code filters before the first question. UUID filenames and vector comparisons do not establish identifier- or embedding-selected push. Unknown external renderer consumers cannot inherit the inspected CLI chains' pull-only inapplicability (RT-RTE-context-query, RT-RTE-ingest-admission).

Source-read log: no source files were read for this profile. All classifications use the supplied accepted records and reconciliation; no new evidence or record is introduced.