pm-llm-wiki memory profile
Type: agentic-system-analyses/types/agentic-system-memory-profile.md
Comparison rationale
The compared memory is the maintained folder, not the external agent runtime. The inventory consists of the source corpus and its digest, product-state pages, optional exploration, catalog, operation journal and changed local instructions. Local files establish the substrate; links do not establish a graph store, optional Git does not establish operative repository-backed memory, and numeric findings do not establish parametric representation. Labels and catalog summaries remain model-interpreted text in the accepted records (MEM-OBJ-sources, MEM-OBJ-index, RT-OBJ-schema).
All positive findings use the claimed basis. This is a prompt-defined design with fictional reported examples, not inspected executing transformation code. Existing files show content availability, not the described agent's execution or reliable compliance. The strongest supported route classifications therefore do not become wired or observed (RT-RTE-source-ingest, RT-RTE-query-fileback).
Reconciliation supersedes the epistemic source, problem, assumption, decision, question, PRD, index, overview and competitor object declarations with their corresponding memory declarations. This profile uses those canonical memory handles. Supersession changes citation identity, not findings: the categorical price generalization and lossy catalog preframing remain limits, while A2's causal qualification remains a supported retained feature. Digest admission and problem confirmation remain distinct parts of ingestion; no route is replaced by either subroute (MEM-OBJ-prd, MEM-OBJ-index, MEM-OBJ-assumptions, MEM-RTE-digest-admission, EPI-RTE-problem-confirmation).
Admission is classified per mechanism. PM curation, digest approval and protected edits are manual even when the model authors and writes them. Ordinary model-admitted wiki changes are automatic even when a PM starts the workflow. New assumption admission and the conflicting evidence-addition clauses prevent complete write-agency coverage. The pilot's confirmed PRD filing is not generalized into a universal human gate (RT-RTE-source-ingest, MEM-RTE-digest-admission, MEM-OBJ-assumptions, RT-RTE-query-fileback).
Knowledge, instruction, ranking, routing and validation name different consumers. Product state supplies evidence and advice; catalog summaries influence relevance; question tags guide proposed discovery; retained decision conditions supply a criterion for reconsideration questions; changed schema supplies standing maintainer directions. These are not executable enforcement. Prescribed supersession withdraws current reliance without deleting history; staleness flags alone do not decay content. The new competitor specification's request to avoid repeated intel does not establish implemented deduplication (MEM-BAP-wiki-synthesis, MEM-OBJ-index, MEM-OBJ-open-questions, MEM-OBJ-decisions, RT-BAP-schema-agent, RT-RTE-stage-upgrade).
Trace learning is not the improved-capacity property. Product-interaction exports supply the approved digest; subsequent automatic wiki synthesis produces durable evidence/problem state intended for later query use. That downstream route supports a claimed yes even though digest admission itself is manual. Its retained input has learning authority at the update and its output has knowledge authority at the later answer. Missing exports prevent a controlled classification of the original trace content. Neither journal retention nor PM-admitted schema progression establishes another automatic trace-learning route; absent qualifying-path evidence is not a bounded negative (MEM-OBJ-digest, MEM-RTE-digest-admission, RT-RTE-source-ingest, MEM-BAP-wiki-synthesis, MEM-OBJ-log, RT-RTE-stage-upgrade).
The read-back inventory is the repository's requested query, ingest, lint and instruction-read chains. Automatic relevance judgment inside a requested read does not create an unsolicited push. Consequently push signal is inapplicable for that complete scoped direction inventory, not an identifier or inferred-judgment finding. External host injection and compaction remain excluded, not established absent (RT-RTE-query-fileback, RT-RTE-source-ingest, RT-RTE-wiki-lint, RT-BAP-schema-agent).
Source-read log: none. Classification used only the supplied accepted members and their contracts; no frozen source ambiguity required a source read.