Echel memory profile

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

Comparison rationale

The two storage families remain separate. Canonical review does not protect the legacy prompt route, and current initialization does not establish a continuous fresh-init-to-agent journey. The comparison retains files and SQLite, rather than treating JSON as a form or relationship edges as a graph substrate. The canonical artifact record is included as access metadata; its externally referenced payload is not classified from that metadata. See MEM-OBJ-canonical-records, MEM-OBJ-disposable-index and RT-RTE-compile-run.

Retained work supplies instructions; knowledge and findings supply context. Canonical strings influence lexical ranking, while status, profile and recovery fields have symbolic routing or gate effects. These are consumer-specific forces, not an implication that every stored statement is accepted or obeyed. Unknown specialized consumers remain included uncertainty. See RT-BAP-context-prompt, MEM-RTE-canonical-query, MEM-RTE-policy-revision and MEM-OBJ-recovery-state.

Human-labelled review is classified only as a local manual decision protocol. It does not authenticate a person or classify generic caller writes as manual. Conversely, automatic retention and import admission do not become manual merely because a caller requested execution or transfer. Direct store writes and caller-dependent initial content keep agency partial. See RT-RTE-knowledge-decide, MEM-OBJ-run and EPI-RTE-portable-import.

Text compilation reduces existing retained material, and the run command retains that representation. This supports consolidation without claiming semantic synthesis. Canonical revisions support evolution; stale/rejected eligibility and resolved finding impact support local invalidation without deleting history. Index rebuilding is a projection, not independently consolidation or promotion. Unrestricted migration permits edits but does not prove a new claim. See RT-OBJ-compiled-context, MEM-OBJ-run, MEM-RTE-staleness-propagation and EPI-RTE-schema-migration.

Requested context selection is pull even though code chooses related IDs. Likewise lexical search on request does not establish inferred-lexical push. Unestablished specialized consumers prevent an axis-wide pull-only conclusion and therefore prevent read-back signal from being wholly inapplicable. Raw stdout/evidence retention establishes trace-extracted lineage, not qualifying trace learning. No positive trace-source finding follows without that qualifying route; nor do the records supply an exhaustive absence warrant. See RT-RTE-compile-run, MEM-RTE-canonical-query, MEM-OBJ-canonical-records, MEM-OBJ-run and MEM-OBJ-evidence.

Reconciliation resolves duplicate epistemic Run, Evidence, clarification, query, finding-review and stale-propagation referents to the corresponding memory records. All citations here use those canonical destinations. Canonical family containers and their epistemic parts, and grouped maintenance/policy routes and their distinct operations, remain nonsuperseded; their scopes are not collapsed. No value is upgraded by supersession. There is no supplied observed execution or causal comparison supporting stronger bases than the recorded wiring.

Source-read log: none. This profile uses only the supplied accepted members and reconciliation; it adds no source evidence or records.