team-ai-workflows memory profile
Type: agentic-system-analyses/types/agentic-system-memory-profile.md
Comparison rationale
The comparison describes the shipped template, not a deployed team's unknown corpus or the host's own memory. Mutable instructions are included only as retained content that team editing can change and later host sessions can consume. The initial static instruction file alone is not evidence of accumulated interaction memory. The file/repository classifications distinguish the local readable payload from its Git distribution surface; access metadata is separate from both. See MEM-OBJ-sources, RT-RTE-repository-admission, RT-OBJ-skill-links and MEM-OBJ-team-workflows-dir.
Reconciliation supersedes the epistemic corpus and search-output identities with MEM-OBJ-sources and MEM-OBJ-search-results. No value changes because of these identity amendments. The latter is temporary output, not another retained store. The container RT-RTE-knowledge-search remains canonical alongside its distinct EPI-RTE-keyword-ranking part. Fixed code using retained corpus words establishes ranking authority; no retained ranking policy is needed.
Requested retrieval is pull even though code automatically scores and slices the response. Full-file reading is a subsequent requested operation, not proven by the preview emission. Host instruction discovery is different: the records establish reachability and an intended later consumer, but not the loader's selection semantics. Accordingly, direction remains partial and push signal remains not determinable, rather than classifying skill-folder names as identifier push. See RT-RTE-knowledge-search, MEM-BAP-corpus-answer and RT-BAP-user-skills.
Manual content editing is an afforded path; mandatory merge review is unresolved. Link admission is wired automatic control, and root-setting admission is automatic only at the claimed setup-procedure layer. Starting a workflow does not approve each resulting value. These layers do not upgrade one another. Editing an existing retained entry affords evolve, but focused-writing advice and preview truncation do not establish consolidation or synthesis of durable memory. No particular adopter transformation licenses a complete curation inventory. See RT-RTE-repository-admission, RT-RTE-skill-sync, RT-RTE-machine-setup and MEM-OBJ-search-results.
The trace-learning negative has a narrower, completely inspected inventory: shipped write and later-consumer routes. MEM-ABS-trace-update covers that inventory, including lookup, synchronization, setup and documented content replacement. Unresolved host delivery semantics do not introduce an unresolved shipped trace-fed writer. The scoped negative therefore survives system union; trace source is inapplicable. This does not deny manual experience-derived tips or classify opaque host/provider learning. It also says nothing about demonstrated improved capacity.
Evidence strength remains local: Python selection and filesystem synchronization are wired, repository customization is afforded, and model-facing directions are claimed. No classification asserts observed operation or causal benefit. Actual configured corpus provenance, merge approval and host discovery/precedence prevent stronger coverage where identified above.
Source-read log: none. This classification used only the supplied accepted members and reconciliation; no source ambiguity required a source read.