Sage-wiki memory profile

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

Comparison rationale

Boundary and coverage

This profile classifies accepted records, not newly inspected implementation. Positive findings describe wired capacity except for external host instruction authority, which is afforded under a conditional loading contract; none establishes observed use or benefit. Every axis is partial because the accepted inventory leaves included alternatives unresolved. The unresolved units are not erased by strong witnesses on other paths. No bounded absence record supports a blanket negative.

The central retained objects are prose and structured knowledge, their access representations, and retained control state. Dense embeddings are parametric representations stored in SQLite, not evidence of updated model weights. Symbolic queue, alias and output-trust fields are included because later invocations consume them to select work, reject candidates or control retrieval. Changed installed guidance is included as retained machinery; shipped static instructions and current-call prompt assembly are not themselves accumulated memory. See MEM-OBJ-vectors, RT-OBJ-compile-items, MEM-OBJ-entity-aliases, MEM-OBJ-pending-outputs and RT-RTE-apply-pack.

Local agency and authority

A caller-authored write does not establish manual agency: the caller could be an agent. Software admission after model generation is automatic even when a human starts compilation. Explicit caller-selected alias/output review and pack/restore selection retain unresolved human control. This uncertainty does not weaken the separately wired automatic branches. See RT-RTE-write-article, RT-RTE-compile-sources, MEM-RTE-resolve-aliases and RT-RTE-admit-output.

Retained prose has knowledge force at later answering/writing consumers; retained text and vectors can also influence ranking. Alias status and output evidence have validation and enforced-veto force at their particular consumers. Those controls are not universal: direct file reads bypass output eligibility. Installed prompt overrides have wired instruction authority at internal completion consumers through registry loading. Generated or installed host guidance instead has afforded instruction authority: writing the files makes standing directions available, but does not wire their force at an external host whose loading, precedence and reload remain unestablished. Neither path proves compliance. The stored learning table has no established native lesson-to-article delivery despite its template slot. See MEM-OBJ-entries, MEM-OBJ-vectors, MEM-OBJ-entity-aliases, RT-RTE-search-wiki, RT-OBJ-prompts, RT-RTE-apply-pack, RT-BAP-generated-skill and MEM-OBJ-learnings.

Read-back chains

Requested search, document answering and graph/community answering are pull operations. Automatic delivery within that requested return is not counted again as push. At a different consumer boundary, compile/resume code automatically selects saved summaries, existing articles and source sections for independent model calls; these are push operations into the receiving completion. Hash/path/concept identity matches and lexical section selection establish different signals; first-section fallback and lifecycle/budget filters are coarse. Retained alias decisions and earlier answers likewise have automatic internal maintenance consumers. See RT-RTE-search-wiki, RT-RTE-query-answer, MEM-RTE-graph-query, MEM-RTE-global-query, MEM-OBJ-summaries, RT-RTE-compile-sources, MEM-RTE-resolve-aliases and EPI-RTE-answer-agreement.

Curation and trace-fed updates

Summary reduction and extraction of standing definitions from recorded session turns support consolidation. Folding near-duplicate concepts, alias consolidation, revision, validity withdrawal, output promotion and age/count pruning have their own local mechanisms. Requested theme/article generation does not by itself establish a new claim absent from the inputs, so synthesis is unresolved rather than silently inferred from a route name. A faithfulness defect also does not turn compression into synthesis. See MEM-OBJ-summaries, RT-RTE-scribe-session, MEM-OBJ-manifest, MEM-RTE-resolve-aliases, EPI-RTE-edge-supersession, EPI-RTE-verify-promote, MEM-OBJ-learnings and EPI-RTE-community-summary.

Scribing establishes trace learning in the profile's write-route sense: recorded user/assistant turns feed automatic durable ontology writes, and later ontology traversal returns their definitions. The trace input has learning authority at the update; the derived definitions have knowledge authority at later consumers. Tool records are filtered out, so this chain supports session-logs, not tool-traces. A pathless entity need not enter ordinary article Q&A. This finding establishes neither improved capacity nor exercised self-improvement. Other capture/source origins and checkpoint/evaluator alternatives remain unresolved rather than inheriting the scribe classification. See RT-RTE-scribe-session, MEM-OBJ-entities and RT-RTE-search-wiki.

Canonical identities and source-read log

Reconciliation supersedes the epistemic raw-source, local graph-answer and global community-answer duplicates with the memory raw-source object and graph/global routes. This profile cites those canonical memory records. The prose-only article, output, relation and community parts remain valid alongside their mixed containers; their distinct consumers are not collapsed. These identity resolutions change citation targets, not classification strength. See MEM-OBJ-raw-sources, MEM-RTE-graph-query, MEM-RTE-global-query, EPI-OBJ-article-text, EPI-OBJ-output-text, EPI-OBJ-relation-state and EPI-OBJ-community-text.

No source files were read for this profile. All warrants come from the supplied accepted records and their annotations; no records, source quotations or new evidence are introduced. Missing batch-payload, initial-provenance, backend/restore-parity, full lint/evaluator-consumer and caller-control facts prevent stronger coverage assessments, not the bounded positive classifications.