browzy.ai memory profile
Type: agentic-system-analyses/types/agentic-system-memory-profile.md
Comparison rationale
The local memory loop separates retained source text, Markdown articles, SQLite search copies and JSON access metadata. Their substrates and consumers are not interchangeable. In particular, temporary raw FTS rows can support human search without a loadable wiki article. See MEM-OBJ-raw, MEM-OBJ-wiki, MEM-OBJ-articles-fts and MEM-OBJ-wiki-index. Both SQLite and its broader relational classification are retained; neither lexical FTS nor numeric backlink scores establishes vector storage or parametric memory.
The schema finding concerns owner changes, not unchanged shipped instructions. Its instruction authority and manual write control are independent of automatic article, digest and draft admission. Preview/skip does not create a per-article human veto. See RT-RTE-schema-change, RT-BAP-schema-instruction, MEM-RTE-interactive-compile and EPI-RTE-article-admission.
Read-back is split at the request and supply operations. Retrieval APIs fulfill requests, while the host selects and supplies articles and recent turns to the answer model. Startup digest selection and post-answer source-prefix selection are separate pushes. Their actual ID matches support identifier selection; requested cache keys alone do not. See RT-RTE-answer-question, RT-RTE-library-query, RT-RTE-session-digest and RT-RTE-insight-draft.
The digest has a retained conversational summary and a later human banner consumer. This supports a wired trace-learning path with knowledge authority at the human display and learning authority at the trace-to-summary update. It does not establish an integrated automatic answer-model memory loop, successful execution or improved capacity. The attempted daily wiki write lacks immediate index admission. Drafts have no established later adoption consumer in the recorded search boundary. The separate compactor return is discarded and is not retained memory. See MEM-OBJ-session-digest, MEM-OBJ-drafts, MEM-RTE-conversation-compaction and RT-RTE-session-digest.
Consolidation does not require that a summary be perfectly faithful. Conversely, a prompt requesting novel connections does not establish implemented synthesis of a new claim. That transformation remains not determinable rather than being classified from its requested behavior. Prefix summaries, replacement writes and forgetting have their own implemented warrants. Index rebuilding, duplicate identification and health suggestions do not by themselves establish deduplication or accepted correction. See EPI-RTE-template-article, RT-RTE-insight-draft, MEM-OBJ-raw and MEM-RTE-health-review.
Reconciliation supersedes the epistemic raw-source and wiki-article objects with the corresponding memory objects, and supersedes the epistemic compaction route with the memory route. This profile uses those canonical declarations; the supersessions change citation identity, not the positive values. The textual image-description part remains distinct from the memory image container. No supersession makes drafts equivalent to digests or search copies equivalent to Markdown. See MEM-OBJ-raw, MEM-OBJ-wiki, MEM-RTE-conversation-compaction, EPI-OBJ-image-description and MEM-OBJ-images.
All ten axes retain incomplete inventory coverage. The accepted memory account leaves profile/streak/history/autocomplete/cost internals unresolved, and the provider component does not reveal use-dependent service state. Source origin and unspecified storage callers also prevent stronger local classifications. These gaps neither erase the inspected routes nor license complete-value statistics. See MEM-OBJ-sessions, MEM-OBJ-activity-log, MEM-OBJ-raw and RT-CMP-llm-provider.
Source-read log: none. Classification used only the supplied accepted records, annotations and reconciliation. No new source passages, records or execution evidence were added.