Project Wiki memory profile

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

Comparison rationale

The inventory follows the memory report's generated artifacts and recorded later consumers, not all storage used by a host coding agent. Repository placement and local file persistence are both supported; Markdown/YAML/JSON describe encoding, not alternative storage services. Original source binaries remain included even though their full representational form and upstream derivation cannot be classified from extracted text (MEM-OBJ-intake-provenance, MEM-OBJ-source-registry).

Reconciliation supersedes the epistemic sidecar duplicate with MEM-OBJ-requirement-evidence; this profile cites the canonical memory record. Chunk text, the ledger and canonical subfamilies remain valid contained parts, not replacements for their broader memory containers. Helper wiring never upgrades semantic integration, instruction installation or future host compliance to observed operation (EPI-RTE-canonical-retain, RT-BAP-project-protocol).

Admission is classified at the decision, not physical I/O or workflow initiation. Normal host integration is automatic in the controlled sense; explicit human supply, preference selection and consequential approval are manual locally. Unspecified direct edits retain unresolved agency. Byte-level quarantine supports deduplication; canonical admission supports a higher tier of use. Requested synthesis does not establish a new claim absent from inputs, so that branch stays unresolved rather than receiving synthesize merely because the workflow calls its output synthesis (EPI-RTE-content-admit, EPI-OBJ-analysis, MEM-OBJ-project-controls).

Automatic preflight and requested file delivery are different operations in one chain. Task-triggered host judgment supplies push selection; page/section/chunk requests are pull. An ID on a requested chunk establishes no identifier-based push. The recorded smallest-relevant-set budget does not establish host eviction, isolated context frames or a token limit (RT-RTE-coding-cycle, MEM-OBJ-wiki-routing, RT-RTE-source-intake).

The chat-answer route supports afforded trace learning because it automatically retains trace-derived material for later task context. At the update consumer, the host consumes conversational experience to decide lasting value and representation and produce durable canonical revisions or analysis; that is afforded learning authority. At the later coding reader, the resulting text supplies knowledge. These distinct consumer/effect paths are supported by the note-update record and its memory annotation, not inferred solely from the trace-learning yes. Neither force requires parameter training, a novel claim or demonstrated benefit. Other write families lack enough recorded trace-origin/qualification detail for a complete inventory. They therefore prevent a complete-value learning conclusion without suppressing the supported yes. No local no is authored: the reports declare no bounded-absence record for these alternatives. Task changes and audit filenames alone cannot establish tool-traces, event-streams or trajectories (RT-RTE-note-update, MEM-OBJ-canonical-records, MEM-OBJ-audit-history).

No source files were read for this mapping; classification uses the supplied accepted records and their annotations. No records or evidence were added. All evidence remains implementation or doctrine/design; no supplied run or experiment warrants activation or benefit.