Arsumbris memory profile
Type: agentic-system-analyses/types/agentic-system-memory-profile.md
Comparison rationale
This is a partial documentary memory inventory, not a classification of the full platform's storage or permission envelope. Files and their in-memory graph are separate operative parts. Indexed assets remain opaque even though their file substrate is known. The ledger's symbolic envelope does not reveal its payload form or backing store. See MEM-OBJ-knowledge-files, MEM-OBJ-live-graph and MEM-OBJ-session-ledger.
The implemented seeder supplies a narrow wired witness for configuration writes and subsequent merge reuse. Host/engine launch consumers remain claimed. Its caller's write opt-in is not per-entry human admission. Parsed rows are preserved, not arbitrary YAML semantics; the parser qualification remains attached to EPI-OBJ-registry-paths. Neither registry merging nor index refresh is semantic curation. The documented overwrite route supports evolving an existing file, without establishing synthesis, truth acceptance or exercised improvement; see RT-RTE-mediated-action.
Explicit file requests and change subscriptions are different operations. A requested path does not support identifier-based push. Subscription details are too thin to determine a push signal. Static skills/injects describe working context rather than accumulated memory; any changed/computed guidance route remains unresolved because its retained inputs, selected payload and later consumption are not established by RT-BAP-profile-guidance. Intended instruction delivery therefore does not create a positive accumulated-memory instruction finding.
Reconciliation supersedes the epistemic knowledge-file, type-definition and session-ledger objects with their memory counterparts, and the epistemic graph-index route with the memory graph-reuse route. All profile citations use those canonical memory identities. The epistemic diagnostic, stamp, registry-part and ingest records remain distinct. These identity resolutions change no local evidence basis and supply no missing operation evidence.
No qualifying automatic trace-to-durable-artifact-to-later-consumer chain is recorded. This prevents a positive trace-learning or trace-source classification, but the unresolved computed-context and package alternatives prevent a bounded negative too. Learning authority is likewise not inferred from event retention. See MEM-OBJ-session-ledger and EPI-RTE-source-ingest.
Source-read log: none. Classification used only the supplied accepted members and reconciliation; no source ambiguity required a direct source read. No execution or causal evidence upgrades the documentary findings.