Signet memory profile

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

Comparison rationale

This profile classifies accepted records only. Every axis is partial because the memory account retains material included alternatives whose objects, admissions, selectors or consumers remain unresolved. Strong local wiring does not resolve those alternatives. No source files were read for this profile; no records or new evidence were added.

Reconciliation asserts no supersessions. Containers and parts remain canonical at their own scopes. The head object, support check and workflow finalization remain distinct: MEM-OBJ-memory-head, EPI-RTE-head-support, and WF-RTE-head-finalization. The corrected head account retains unchanged-current-body and already-empty-head NOOP branches before submitted support validation. Enforcing validation applies at its actual checks, not to every successful finalization. Database revision, file publication and conditional startup supply are separate outcomes.

Graph and SQLite findings describe access organization and physical backing of the same retained graph parts. Dense embeddings are parametric representations, not proof of provider-weight learning. Localized feedback weights are symbolic because their consumers use fixed update and ordering rules. Transient in-memory prompts and staged submissions are not independently accumulated objects. See MEM-OBJ-entity-attributes, MEM-CMP-embedding-provider, and MEM-OBJ-path-feedback.

Admission is not authorship. Software/model gates establish automatic writes for named transformations; an arbitrary caller or an authored filename does not establish a human per-content decision. Initial identity approval, rating authorship and other import/recovery controls remain unresolved. See RT-RTE-direct-remember, MEM-OBJ-identity-files, and MEM-RTE-path-feedback.

History leaf extraction and parent compression support consolidation without requiring semantic fidelity or hard byte ceilings. Aggregate duplicate handling and refresh support deduplication and evolution. Forgetting and tombstone/staleness transitions support forgetting and invalidation at current-reliance consumers, not privacy erasure. The route called synthesis does not establish a new claim absent from its inputs: the accepted epistemic account leaves that relation indeterminate. Requested head condensation is kept at claimed basis, separate from wired history reduction. See MEM-RTE-history-compaction, MEM-RTE-state-withdrawal, EPI-RTE-aggregate-synthesis, and WF-RTE-head-finalization.

Explicit recall and zoom are pull. Automatically supplied startup background and initial maintenance history are push, even when a lifecycle request precedes them. Scope matching actually selects history streams; filenames on requested reads do not establish identifier push. Recall's lexical/vector capabilities do not by themselves prove the startup selector's signal. Inherited, recovery and predicted-context selectors and other host adapters remain unresolved. See MEM-RTE-explicit-recall, RT-RTE-pi-lifecycle, and MEM-RTE-history-compaction.

Trace learning is established by automatic session-fed semantic/head writes, pass/tool-fed history summaries and retrieval-feedback-fed selector weights with later consumers. Their update inputs have learning authority; outputs independently have knowledge or ranking authority. Host compaction supplies another automatically retained continuation-summary route, without resolving its complete original inputs. This does not establish exercised improvement. Host summary origins and aggregate mixed-evidence origins remain locally unresolved. Tool activity is not automatically a multi-step trajectory, and conversational material wrapped in capture events remains session logs. See MEM-RTE-transcript-capture, RT-RTE-dreaming-pass, MEM-RTE-history-compaction, MEM-RTE-path-feedback, and MEM-OBJ-session-summaries.

Benchmark checkpoint status routes subsequent evaluation, but retained scores are not established product-memory acceptance or optimization inputs. Unknown serialization and unread optimization consumers prevent stronger substrate, lineage and trace conclusions. See MEM-OBJ-run-checkpoint and EPI-RTE-benchmark-checkpoint. No bounded absence records support a system-wide negative on any axis.