Gnosis memory profile

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

Comparison rationale

The comparison boundary is accumulated entries and their retained access state, not the static help text or the external coding runtime. The four operative parts are MEM-OBJ-entry-text, MEM-OBJ-entry-metadata, MEM-OBJ-entries-fts and MEM-OBJ-metadata. Decoded objects, snippets and topic counts are consumption stages, not independently accumulated stores. Repository sharing is an afforded use of the canonical file, not a wired replication service.

Reconciliation supersedes the epistemic declarations of entry text, entry metadata and entries FTS with the corresponding memory declarations. This profile cites those canonical memory objects. The epistemic import, check, replacement, withdrawal and refresh routes remain distinct functional parts of valid runtime containers; they are not superseded. These connections change citation ownership, not evidence strength.

Authority is consumer-specific. Entry prose provides advisory knowledge at the external agent, while retained text/topics influence lexical priority and metadata controls lookup and refresh. Existing IDs also act as criteria and veto inputs for related-reference admission. These are structural validation and enforcement, not epistemic acceptance of body claims. See RT-BAP-agent-guidance, EPI-RTE-structure-check and EPI-RTE-edit-check. Unauthenticated author metadata and displayed Related pointers do not confer additional force.

Editing supports evolution; deletion supports forgetting. Neither establishes history-preserving invalidation. Writing a related replacement does not automatically suppress the original, and cache freshness is not reliable semantic withdrawal. The limits follow EPI-RTE-entry-replacement, EPI-RTE-entry-withdrawal and MEM-OBJ-metadata. No consolidation or synthesis finding is inferred from a request to review session knowledge; the original session and the caller's actual transformation are unresolved.

All recorded read-back routes fulfill explicit requests. Ranking and identifier lookup inside those requests are not push delivery, so the push-signal axis is inapplicable even though lexical and identity selectors exist. This conclusion uses the complete shipped read-route inventory and annotation on RT-RTE-knowledge-read, not a claim that every possible external integration was inspected.

The automatic generation of fields and derived indexes does not resolve who decided to admit prose. Likewise, an afforded session-extraction path does not establish an automatic trace-fed write. Unknown per-content caller control and original inputs remain in both trace axes. No positive trace-learning value or system-wide negative is asserted. No source files were read for this classification; all warrants and coverage limits come from the supplied accepted records. No operation, compliance, improved capacity or causal benefit is inferred from implementation evidence.