LLM Wiki v2 memory profile
Type: agentic-system-analyses/types/agentic-system-memory-profile.md
Comparison rationale
This profile classifies accepted records, not new source evidence. Reconciliation establishes no whole-record supersessions. The source/candidate containers remain distinct from retained raw and note parts; classification uses the latter wherever their consumption differs. The unresolved check-result part separation prevents a complete evaluator-level classification. It does not negate the separately recorded retention and offline consumption in MEM-OBJ-evaluation-results.
Storage is not model architecture. Notes, indexes and provenance are file-backed; the classifier response cache is SQLite. Calling a distributed-parametric classifier does not make its retained scalar answers parametric memory. Neither file retention nor imported-source lineage establishes truth warrant. The readable answer display carries title, summary and main body, not every field or linked target: MEM-OBJ-note-body, MEM-OBJ-note-metadata, MEM-OBJ-answer-cache.
Admission is not authorship. Default model/software transformation is automatic even when a human starts ingest. Caller-provided edits, proposal origins and confirmation flags do not establish human content approval. The profile therefore retains those mechanisms as unresolved rather than inventing a manual witness: RT-RTE-source-filing, RT-RTE-manual-maintenance, MEM-OBJ-proposals.
Authority follows the actual consumer. Prose supplies knowledge to later writing and answering. Summaries/descriptions influence ranking and routing; fields, hashes and queue status can supply validation and enforced vetoes. These do not give arbitrary instructions embedded in notes standing authority. Automatic archive replacement withdraws ordinary active reliance while retaining a historical read route. Purge instead deletes referenced lineage and refiles other sources; source-removal marking leaves derived notes usable. Those distinctions prevent a blanket invalidation classification: MEM-BAP-retained-knowledge, MEM-OBJ-archive, EPI-RTE-origin-purge, MEM-OBJ-raw.
Read-back chains contain different operations. Asking a question, requesting a note or requesting a model response is pull. Automatic supply of retained context to internal filing consumers and folder-specific prior-attempt state to splitting is push. The folder identity match is a targeted signal because it actually selects the supplied count; a note path on a requested read alone is not a push signal. This distinction does not establish proactive injection into an enclosing client: RT-RTE-question-answer, MEM-RTE-note-read, MEM-RTE-answer-reuse, MEM-OBJ-log.
The trace-learning positive is deliberately local. Split attempts feed an automatically retained control artifact, and later splitting consumes it to change retry behavior. The original input is operational events, not session logs. This supports wired trace learning and learning authority at that update path, not empirical improvement. Memoized model results and sync checkpoints are recorded automatic control updates, but the records do not settle every qualifying behavior-shaping update and original input there. Client-written durable experiences also leave extraction control unresolved. These branches remain separate unknowns, not negative findings: MEM-OBJ-log, RT-RTE-batch-sync, MEM-RTE-answer-reuse, MEM-OBJ-sync-state, MEM-OBJ-raw.
All positive findings are implementation-level wiring. No finding is promoted to observed operation or causal benefit. Incomplete custom implementation, caller provenance and evaluator coverage keep the relevant inventories partial. No bounded-absence record is declared in the supplied members, so the profile supplies neither a local trace-learning negative nor a system-wide negative.
Source-read log: none. This mapping used only the supplied accepted members and governing contracts; no frozen-source ambiguity required a source reread.