Sparks memory profile

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

Comparison rationale

This profile classifies the accepted records, not a running personal vault. Wiring establishes deterministic file/SQLite writes and selection. The external agent's semantic obligations remain claimed, and its later body-reading route is afforded. No classification is observed or causally supported: the records supply no execution trace or comparison. The scope separates accumulated content from the static RT-OBJ-contract and the external RT-CMP-agent; their persistence or parametric computation does not turn them into accumulated Sparks vault memory.

Inventory and local uncertainty

The raw, wiki, log, Tasks, index, generated-collection, manifest-projection and ingest-history units retain distinct consumption paths. Markdown has natural-language bodies and symbolic dates, links, fields and markers. The MEM-OBJ-manifest-projections is executable access state rather than source-body meaning. Optional repository history follows RT-RTE-ingest-cycle, not an assertion that every vault has Git enabled. Available helper details and absent executions limit extraction quality or activation; they do not make the inspected substrate opaque.

Known deterministic admission does not settle initial provenance or generic caller identity. In particular, MEM-OBJ-tasks has software target admission and unresolved caller-supplied content control. The MEM-OBJ-index preserves old descriptions without authenticating their initial author or reevaluating them. Those unresolved alternatives remain in lineage and write-agency units. The prescribed external-agent writer has automatic admission because no per-page human decision is required, not because a human started the ingest.

Authority, curation and read-back

Page and log content supplies knowledge at readers; it is not standing host instruction. Manifest/session state supplies routing and lifecycle admission criteria. Retained task text influences fuzzy ranking, while checkbox state controls the candidate/preview set. These consumer-specific effects do not depend on demonstrated answer improvement. Task identity matching and row interlocks also do not license semantic truth or real-world task success: see RT-RTE-maintenance-write, RT-OBJ-ingests and RT-RTE-vault-consume.

The two wired synthesis findings are local: EPI-RTE-books-extract infers probable book identity from attribution punctuation, and EPI-RTE-projects-extract turns maturity into shipping labels. Their accepted non-entailment warrants distinguish these from literal hint import, source-field selection and context reduction. EPI-RTE-reading-extract implements exact-text deduplication. Requested semantic summaries do not establish synthesis, and uncertain Bookmarks/Media associations remain unresolved rather than inheriting a neighboring extractor's basis. Likewise generated stable headers do not establish promotion through review.

Requested collection regeneration consumes retained sources and produces files for a later requested reader. Query supplies metadata, not Revision reasons; the prescribed full-body read is the bridge to those reasons. Brief delivers selected retained log/task text without automatically saving a new summary. The RT-RTE-vault-consume and MEM-RTE-collection-regeneration therefore support pull, not a second push merely because fixed code selects and delivers material. Read-back signal is inapplicable for this fully named pull-only vault inventory; no claim is made about the excluded harness loader.

Trace-learning remains not-determinable rather than a controlled negative. The MEM-OBJ-log is requested authored operational history; its forwarding to Brief is not a recorded automatic trace-fed durable update. Archive and access-state derivation alone do not establish that chain either. The accepted records declare no bounded absence supplying a local no. Consequently trace-source cannot be made inapplicable by an unsupported system negative, and dated captures cannot be relabeled event-stream inputs without a qualifying write.

Reconciliation and source-read log

The supplied reconciliation supersedes the epistemic raw object with MEM-OBJ-raw, the epistemic task object with MEM-OBJ-tasks, and the epistemic affected route with MEM-RTE-affected-selection. All values above use those canonical destinations. The combined epistemic wiki object and seven-output memory collection container remain valid alongside their parts; they are not reasons to merge consumer effects or strengthen one branch from another's evidence.

The recorded missing-frontmatter conflict prevents assuming tolerated scan files remain ordinary query candidates. No value here depends on that implication. Missing-file withdrawal is limited to catalog visibility, not complete graph withdrawal or truth-based retraction: MEM-OBJ-manifest-projections retain the stale-edge qualification. Thus neither catalog routing nor bounded invalidation promises complete freshness.

Source-read log: none. This job used only the supplied accepted members and governing contracts; it added no source passages, records or evidence and did not reopen report verification.