Napkin memory profile

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

Comparison rationale

The scope is accumulated vault material and its access/update routes, not every storage object in the runtime. Temporary benchmark vaults count because later question invocations consume earlier driver writes; transient SQLite query databases and reconstructed link/graph snapshots are not persistent memory substrates. External model weights are components outside this vault-memory boundary, not evidence for model-weight storage or parametric vault memory. See MEM-OBJ-base, MEM-RTE-benchmark-variants, MEM-OBJ-search-cache and RT-CMP-pi-model.

No supersession affects a value. Reconciliation retains valid containers and their parts. This profile uses the memory object's finer distinctions without replacing the broader epistemic containers. In particular, the supplied LoCoMo summary's inspected copy supports imported lineage, not a Napkin extraction route or its unknown upstream learning mechanism. See MEM-OBJ-session-summary and EPI-OBJ-benchmark-notes.

Requested semantic changes remain claimed. The packaged session procedure and separate-model background design each describe automatic session-fed durable writes with later consumers, so each supports claimed trace learning. They are distinct routes, not one wired scheduler. This classification neither establishes learning benefit nor settles whether a concrete inferred claim was novel or warranted. The post-write diagnostic checks links, not admission of knowledge. Template/path mismatches further limit successful delivery of the requested distillation. See RT-RTE-distill-session, MEM-RTE-background-distill, WF-RTE-distill-verification and RT-RTE-caller-mutation.

Consolidation follows the specified reduction of session material to selected topic-based conclusions; it does not depend on proving that every resulting sentence is faithful. Deduplication and evolution follow their own requested merge/rewrite mechanisms. Inference markers alone do not establish synthesis of a claim absent from the inputs. Cache derivation and index rebuilding alone are not semantic curation. Trash invalidates ordinary search/listing reliance but leaves exact-path read capacity; it is not global access enforcement. The separate permanent-delete branch forgets the retained note through unlink and therefore supplies wired decay, independently of timed or scored decay policies. This resolved branch is included in the operation inventory even though other transformations remain unresolved. See RT-RTE-distill-session, RT-RTE-tend-vault, EPI-OBJ-vault-prose, MEM-OBJ-search-cache and MEM-OBJ-markdown-notes.

Knowledge, ranking and routing are separate effects. Retained note terms affect fixed lexical ranking even without a learned ranking policy. Templates and conventions afford guidance, while their delivery and interpretation by a host remain uncertain. Static skill instructions, executable create guards and result-score heuristics are not automatically retained-memory validation/enforcement criteria. Evaluation reports affect aggregation, not an inspected durable policy update. See RT-RTE-vault-retrieval, MEM-OBJ-note-templates, MEM-OBJ-napkin and EPI-RTE-score-report.

Core read-back fulfills explicit requests. Bundling NAPKIN.md with requested overview does not create a second push. Its always-loaded host description supports only claimed broad/coarse push: external scheduling, identity matching and injection remain unresolved. Requested filenames and base/view selectors cannot establish identifier-based push, and lexical search cannot establish lexical push. The opaque alternative therefore keeps direction and signal coverage partial. See RT-RTE-vault-retrieval, MEM-RTE-structured-access, MEM-OBJ-napkin and MEM-RTE-background-distill.

Raw dialogue ingestion alone is not trace learning. Supplied summaries have unknown production; stored scores have no inspected successor-policy consumer. Those limits are not bounded absence records and cannot justify a system-wide no or inapplicable trace source. Known session-message origin does not turn timer triggers into event-stream inputs, nor prove that findings include original tool traces or trajectories. See RT-RTE-benchmark-answer, MEM-OBJ-session-summary, MEM-OBJ-evaluation-results and MEM-RTE-background-distill.

Source-read log: none. Classification used only the supplied accepted members and reconciliation; no frozen source file was read and no new evidence was introduced. Missing initial admission/provenance, opaque index semantics, external cursor/injection and partly inspected query/rendering consumers remain local limits rather than being erased by stronger witnesses elsewhere.