ReframeWeb memory profile

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

Comparison rationale

Boundary and evidence strength

The profile classifies retained operative parts, not every place the enclosing runtime stores a value. Graph organization and local KV backing coexist; checkpoint bytes have SQLite metadata and filesystem payloads. Locator display prose does not classify file bytes, and symbolic image descriptors do not classify the rendered scene. See MEM-OBJ-context-memory, MEM-OBJ-workspace-store, and EPI-OBJ-rendered-evidence.

The supplied reconciliation makes no supersession. Containers and their declared parts therefore retain separate canonical citations. Findings use implementation warrants except the standalone import's afforded invocation. No observation or causal experiment upgrades the mechanisms to exercised operation, reliable compliance, factual acceptance or beneficial learning.

Admission, transformation and force

Human exact-content browser edits support manual agency; unspecified API/script callers do not. Automatic message/receipt generation and computational guidance review retain their independent positive warrants. Initial preference or instruction labels do not settle actual content admission. See MEM-RTE-node-curation, RT-RTE-guidance-review, and RT-RTE-application-publish.

The summary is consolidation of retained action history, not demonstrated novel knowledge. Guidance review's request for distinct durable lessons is not an implemented duplicate merge or evidence of synthesis. Graph deletion supports forgetting, while uninspected blob cleanup prevents a complete lifecycle inventory. Retained tool history has learning force at the summarizer that produces the later-invocation completion/retry summary; the summary has knowledge force at those later consumers. Guidance proposals, by contrast, are transient staged candidates, not established read-back of retained decision observations. The review also receives stored guidance, but the records do not establish a separate learning force for that input. The guidance route's architectural trace-learning positive and its output's later instruction force do not supply that missing input-authority warrant. None implies observed improved capacity. See MEM-OBJ-task-history, MEM-RTE-history-summary, EPI-OBJ-candidate-entry, EPI-RTE-guidance-check, and MEM-RTE-node-curation.

Saved controls can route, validate and enforce. Their effects do not come from the instruction prose. Current expected text vetoes a conflicting replacement; saved task grants and activation bindings govern covered operations. No universal effect monitor is inferred. See RT-RTE-saved-store, RT-OBJ-tool-attempt, and MEM-BAP-linked-instructions.

Requested retrieval and automatic supply

Requested graph/history retrieval is pull. Automatic context reload, model relevance selection, explicit-link materialization and next-inference image attachment are push operations within their respective chains. Lexical search hints establish the requested retrieval selector, not a separate inferred-lexical push. The model's kept IDs represent semantic judgment; explicit task/Store identity matches instead select associated retained parts without that relevance decision. See MEM-RTE-graph-selection, MEM-RTE-context-reload, and EPI-RTE-panel-evidence.

Trace qualification and unresolved origin

The optional observation-to-guidance write qualifies architecturally. Its review interface does not establish the original trace's controlled source category. Candidate narration cannot replace a missing original-input fact. The action-summary route has a recorded tool-history input and a later completion/retry consumer. Under the shared contract, replacement summaries retained for a later invocation are memory even within one request; this positive is bounded to that horizon and does not claim cross-session summary retrieval. See RT-RTE-guidance-review, EPI-RTE-guidance-check, and MEM-RTE-history-summary.

Unknown initialization, arbitrary payload forms, generic-caller control and uninspected maintenance/later-consumer alternatives remain in the axis inventories. They cannot be dropped to turn a wired witness into complete coverage. No local or system-wide trace-learning negative is asserted without a bounded absence record; therefore trace_source is not made inapplicable by raw-log or evidence-only routes.

Source-read log

No frozen-source file was read for this mapping. All classifications derive from the supplied accepted records and their annotations. No records, quotations or source evidence were added, and report verification was not reopened.