OrangePro memory profile

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

Comparison rationale

The comparison follows the accepted records and their annotations. Reconciliation establishes no duplicate supersessions; valid containers and parts remain distinct. Every axis is partial because local witnesses do not enumerate opaque hosted state or all included local alternatives. Uninspected helpers and unknown admission provenance remain uncertainty, not absence.

Filesystem JSON establishes files, not a graph database substrate. Symbolic graph structure and executable code are separated from interpreted descriptions and reasons. External model parameters are inference components rather than an OrangePro use-updated memory object; their opacity does not turn local symbolic scores into parametric memory. See RT-OBJ-graph, RT-OBJ-generated-tests and RT-RTE-platform-call.

The outer request can pull a saved product while internal selection pushes retained context into a later provider invocation. Target/relationship matches select graph evidence; the provider did not independently request those retained parts. Budget filtering constrains delivery but does not establish semantic similarity. See MEM-RTE-context-selection and MEM-RTE-start-reuse.

Trace learning here uses the profile's automatic trace-fed durable-artifact criterion. A repaired test and a retained proof decision have later consumers, although neither establishes improved capacity, a revised general strategy or weight training. The proof record's downstream force is routing/validation rather than a blanket instruction to a model. Repair feedback has learning authority at the update consumer and produces executable instructions for later verification. See RT-RTE-draft-verification, MEM-OBJ-ledger, RT-BAP-proof-credit and EPI-RTE-export-replay.

No evidenced-absence records were declared in the supplied reports. Non-qualifying recomputation/display examples therefore do not warrant local trace-learning no findings or inapplicable trace sources across unresolved alternatives. The supported yes findings survive the partial inventory. Generic caller confirmations likewise do not establish manual content admission; Actions may supply them automatically. See RT-RTE-generation-compare, RT-RTE-actions-delivery and MEM-OBJ-parse-cache.

Curation follows transformations rather than labels: candidate deduplication, stale-mark withdrawal, setup repair and distilled diagnostics have separate warrants. Numeric quality proxies do not establish synthesis or semantic acceptance. Proof credit is a target tier, not by itself promotion of a retained memory entry. See RT-RTE-candidate-admission, RT-OBJ-generated-tests, WF-RTE-proof-recovery and EPI-RTE-comparison-scoring.

All bases describe implementation wiring, not observed activation or causal benefit. No frozen-source reads were needed or performed for this classification; the supplied accepted records provide its evidence passages.