QwenPaw memory profile

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

Comparison rationale

All ten axes are partial. Local witnesses establish real classifications, but the accepted records do not supply a complete inventory of provider-owned memory, initial contents or auxiliary stores. Unread available implementations and unavailable external internals are different limits. No absence record supports a system-wide negative, and no run or causal evidence upgrades a wired route to observed operation or benefit.

The profile follows the reconciliation's identity amendments. It cites canonical MEM-RTE-trace-extraction, MEM-RTE-dream-maintenance, MEM-RTE-paper-maintenance, MEM-RTE-memory-edit and MEM-CMP-reranker-model, rather than their superseded epistemic declarations. Their epistemic qualifications survive: dispatch/completion is not factual acceptance, dream completion may follow a null response, and backend semantic operations remain opaque. Non-superseded relevance subfunctions remain separate witnesses.

The session container is not replaced by its summary/index parts. RT-OBJ-session-state, MEM-OBJ-continuation-summary and MEM-OBJ-eviction-index classify different consumed parts. The continuation summary supplies context; its source ranges and recovery map additionally route and constrain recovery. A user-role background message is not automatically a standing instruction. Actual system-prompt authority instead follows RT-BAP-workspace-prompt.

Trace learning is positive for automatic derived notes and continuation state with later consumers, independently of semantic novelty, parameter updates or successful improvement. Retained experience has learning force at the update consumer, while produced text has knowledge force at later task consumers. MEM-RTE-scroll-retention supports implemented consolidation, replacement and provenance-based withdrawal. ReMe's described fact extraction and dream merging retain claimed curation findings, not an upgraded wired semantic transformation. Neither opaque analysis nor a potentially inaccurate summary establishes synthesis.

Retrieval direction follows operations, not delivery mechanics. MEM-RTE-memory-retrieval supports requested pull and middleware push. Keyword/vector access and optional judgment-based reranking concern automatic query selection; the explicit search return does not become another push. MEM-RTE-history-recall remains pull even when parameters are IDs. Conversely session restore and configured prompt contribution use actual identity matches selecting delivered retained parts. Availability and summary-render budgets are coarse filtering, not semantic targeting. Proactive selection remains partial because its exact recent-session predicate is unresolved.

WF-RTE-fork-finalization supports automatic Git/registry admission and administrative validation/enforcement. It does not establish product correctness or learning from the product. The workflow's saved-parent-file inheritance narrows fork context selection: it may be stale or empty. External harness resume establishes an outer identity-selection path, not complete provider memory semantics. Neither project retention nor model inference admits all runtime storage into the comparison boundary.

No additional source files were read. Classifications use only the supplied accepted records and their retained evidence. Unknown admission control for generic caller replacements is preserved rather than converted into manual agency; starting extraction or restoration is not per-content human approval. Remote returned text is classified separately from opaque stored payloads, and inbox display text never substitutes for the underlying modified memory artifact.