BRAN memory profile

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

Comparison rationale

The comparison follows operative retained parts and their actual consumers, not the runtime's complete storage surface. Repository evidence, query-use history, local results and settings have different write and read routes. Optional file-state, custom-result-store and derived-file APIs remain visible with capability-level findings or unresolved units; they are not silently promoted to integrated agent memory. Immediate database extraction and Obsidian export/reparse establish no retained later-consumer route, so their temporary buffers add neither an in-memory substrate nor learning findings. See MEM-CMP-database-evidence, MEM-CMP-obsidian-export, MEM-OBJ-file-state and MEM-RTE-derived-rebuild.

The supplied reconciliation establishes no duplicate supersession. All citations therefore retain their declaring members and canonical IDs. Document and SCIP routes refine the query/packet container; diagnostic summary/replay routes are downstream consumers within the broader feedback chain. Provider claim candidates and retained InlineResult support are different stages, not interchangeable object declarations. MEM-RTE-document-selection, MEM-RTE-symbol-reload, MEM-RTE-evidence-feedback, EPI-RTE-evidence-summary, EPI-RTE-evidence-replay, EPI-OBJ-provider-claims and MEM-OBJ-result-support keep those distinctions.

Retention, form and provenance

Files are the supported retained substrate, including opaque result artifacts whose storage is known even though their meaning is not. The custom runtime store port remains unresolved. A versioned local record is still file storage; neither content-addressing nor numeric digests imply a vector store or parametric representation. RT-RTE-runtime-port does not identify a concrete alternative store.

Natural-language answers, request text and repository prose are split from symbolic access descriptors, source associations and executable settings. Snapshot/validator framing does not resolve enclosed bytes. The display report's truncation does not classify the raw snapshot or generated-validator payload. Deterministic canonical encoding supports other-compiled lineage only for that encoding step; it does not settle original provider derivation. Initial repository authorship and arbitrary caller payload origins remain unresolved. See MEM-OBJ-derived-report, MEM-OBJ-derived-snapshot, MEM-OBJ-generated-validator and MEM-OBJ-lossless-artifact.

Authority and admission

Knowledge supply and ranking are separate consumers of repository material. At grounded completion, retained source bytes additionally supply validation criteria and an enforced veto, without becoming a truth oracle. Query history supplies diagnostic evidence and selects replay inputs; it does not automatically revise ranking policy. Settings route later invocations and constrain budgets through executable consumers. These effects do not establish standing natural-language instructions or learning authority for arbitrary stored bytes. RT-RTE-query-packet, RT-RTE-connected-request, MEM-RTE-evidence-feedback and RT-BAP-settings-consumption support the separate paths.

TUI Apply establishes manual admission for that UI route. Software query/result admission and the derived-only helper establish automatic mechanisms at their respective evidence layers. A generic caller and an asserted maintenance authority do not authenticate human review. Lower-level settings and initial content therefore retain uncertainty independently of the inspected TUI choice. The helper's afforded automatic authority does not override the CLI fixture gate. See RT-RTE-settings-apply, RT-RTE-maintain-apply and MEM-RTE-derived-rebuild.

Replacement supports evolve and result eviction supports decay. Exact identity collisions, canonical encoding and index rebuilding alone do not establish dedup or consolidation. Query clear is removal, not a demonstrated history-preserving invalidate operation. No controlled synthesis value is inferred from advisory proposals whose replacement and authority are null. See MEM-RTE-evidence-feedback and the memory annotation on RT-RTE-connected-request.

Read-back and learning limits

Explicit query/packet, history commands and get are pulls, including identity-targeted get. They do not become push merely because code delivers the response. Automatic settings loading is coarse push: setup selects the available project configuration without a content-relevance selector. Grounded completion separately selects cited files by locator for validation, an identifier-based automatic supply to its checker. Query's lexical ranking is not a push signal on a requested evidence return. Opaque artifact integrity read-back immediately after publication does not establish a later consumer invocation. These distinctions follow RT-RTE-query-packet, RT-RTE-connected-request, RT-RTE-settings-apply and MEM-RTE-evidence-feedback.

Trace-extracted query descriptors do not by themselves establish trace learning. The recorded chain ends in advisory diagnostics and replay rather than an established automatic durable behavior-shaping revision. Connected output retention likewise does not establish a trace-fed update, while arbitrary artifacts, derived inputs and custom ports leave provenance and later consumption unresolved. No supplied member declares an evidenced-absence record, so no local or system-wide trace-learning "no" is warranted. Trace source is not-determinable, not inapplicable; no bounded complete absence resolves its applicability. These limits preserve the positive retention and diagnostic mechanisms without asserting improved capacity or successful self-improvement. See MEM-RTE-evidence-feedback, MEM-RTE-derived-rebuild, MEM-OBJ-lossless-artifact and RT-RTE-runtime-port.

All positive findings retain wired or afforded evidence at the actual mechanism. No observed execution or causal support is supplied; a wired local route does not resolve an opaque alternative or demonstrate activation or benefit.

Source-read log

No source checkout files were read. Classification used only the supplied accepted members and contracts; no named record ambiguity required direct source inspection.