WikiBrain memory profile

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

Comparison rationale

This profile classifies the accepted records, not a running service. All positive findings are implementation wiring unless explicitly marked claimed. None is an observation of model compliance, successful retrieval, semantic correctness or improved capacity. The ordinary compilation route admits individual page changes before any mandatory semantic acceptance; the retained wiki therefore supplies knowledge/context without thereby becoming warranted truth. See RT-RTE-ingest-job and MEM-OBJ-wiki.

Canonical identities. Reconciliation supersedes the epistemic raw and wiki object declarations with MEM-OBJ-raw and MEM-OBJ-wiki. This profile uses those canonical objects. Import-fidelity limits remain supported by EPI-RTE-source-import. The structural lint return is not the retained wiki lint report; live schema is not its delivered template base or custom bundle. These distinctions prevent spurious in-memory storage and premature instruction authority for upstream text. See EPI-OBJ-lint-report, MEM-OBJ-schema, MEM-OBJ-template-state, and MEM-OBJ-user-templates.

Substrate and form. Markdown paths, linked pages and exported assets do not make the operative store files, repo or graph: the accepted objects are PostgreSQL rows/JSON/bytea, including the relational link table. Natural-language bodies and code-consumed selectors remain separate forms. Image payloads are not distributed parameters, and their visual representation lacks a supported controlled classification. The unresolved visual unit makes form partial without making PostgreSQL storage uncertain. See MEM-OBJ-links-tags and MEM-OBJ-assets.

Admission and curation. A human-started job can admit automatic model writes. Browser selection/editing supplies a narrower manual witness; generic write-token identity does not settle per-content control. The alternate normalized schema path is another automatic admission, not evidence of actual self-modification in a run. See RT-RTE-schema-spelling, RT-RTE-client-tools, and RT-RTE-rule-update. Revision, soft withdrawal and snapshot expiry support evolve, invalidate and decay. Requested summaries support only claimed consolidation. The semantic relation of generated content remains unresolved, so a requested synthesis or lint diagnosis is not upgraded to implemented semantic synthesis or demonstrated criticism. See MEM-RTE-state-withdrawal, EPI-RTE-semantic-check, and RT-RTE-ingest-job.

Read-back operations. A requested get_instructions response is pull even though it includes automatically chosen pending hints. By contrast, next-turn history is automatically supplied after the session is selected, and default compilation receives code-selected pending targets without the model first requesting that retained set. The pending selector is coarse availability/backlink gating; the conversation selector actually matches a session identity. Lexical search supplies no inferred-lexical push finding because it answers a request. See MEM-RTE-note-retrieval, RT-RTE-chat-session, and EPI-RTE-pending-selection.

Trace-fed updates. Raw log retention alone is insufficient. The positive chains are more specific: retained discussion turns feed automatically admitted wiki text with a later read route; tool events feed diagnostic written-path fallback that can be retained as assistant prose and supplied in a later turn. Operator initiation of discussion compilation is not approval of each output page. These wired chains establish local trace-learning yes without asserting improved future capacity. Their original inputs support session-logs and tool-traces respectively; the job event wrapper does not establish environmental event-stream origin. See MEM-RTE-discussion-guidance, RT-OBJ-ingest-jobs, MEM-OBJ-chat-sessions, and RT-RTE-chat-session.

Ordinary history retention without those derived branches and opaque external-content admissions remain unresolved learning units. No accepted evidenced-absence record supports a local no. The derived system union therefore retains yes and the known trace sources but cannot claim complete coverage or any system-wide no. The discussion transcript has learning authority at the automatic wiki-writing consumer, not at ordinary history projection. Copying saved role/content into a later history supplies knowledge/context; it does not establish an update to parameters or a durable behavior-shaping artifact at that projection consumer. The resulting wiki/fallback prose supplies knowledge at later readers. Requested log/reason retention does not guarantee a later reader receives all reasons. See MEM-OBJ-wiki and MEM-RTE-discussion-guidance.

Coverage limits and source reads. Scope concerns the shipped local retained loop, not provider internals or private MCP-client policies. Generic caller provenance, qualifying ordinary-history derivation, visual payload form and actual model transformations prevent stronger assessments where named. Uninspected template-language content does not make its inspected PostgreSQL storage or rule-delivery path unknown, but it prevents classifying every template's semantics or epistemic method. No additional source reads were needed or performed; all warrants remain in the supplied accepted records. This profile declares no records and adds no evidence.