Amanuensis memory profile

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

Comparison rationale

This profile classifies accepted records, not a new source inspection. All positive implementation findings remain wired; model-facing survey and standing guidance retain afforded status. No finding is upgraded to observed operation or causal benefit.

Boundary and canonical identities

The retained unit is the conspectus and its later-consumer routes, not the whole host's storage. Static phase references and MCP opening guidance are excluded from accumulated memory; the changing standing delivered to an answering host remains included through MEM-BAP-recorded-standing. External provider opacity does not add model-weights or parametric memory values. RT-RTE-survey-cycle bounds host orchestration, persistence and resume; RT-RTE-tool-dispatch bounds the service rather than every alternate file writer.

Reconciliation supersedes the epistemic claims, findings, explanations, research-results and stale-ledger objects with the corresponding memory declarations. All classifications therefore cite MEM-OBJ-claims, MEM-OBJ-findings, MEM-OBJ-explanations, MEM-OBJ-research-results and MEM-OBJ-file-ledger, not their superseded duplicates. The narrower epistemic routes remain canonical and usable; no route is replaced merely because it overlaps a broader memory lifecycle. These supersessions affect citation identity, not controlled values or evidence strength.

Content, metadata and force

Natural-language statements and explanations are split from symbolic keys, intervals and dispositions. Both are retained in SQLite, but consumers differ: a model interprets content as knowledge; fixed gates use evidence links and slot state as validation criteria and enforce refusals. EPI-RTE-explanation-selection illustrates why a retained positive body need not retain current authority after a decline or withdrawal. The status-admission exception in RT-RTE-survey-cycle prevents turning these local gates into universal completion enforcement.

Ranking includes retained candidate text, demand and heat even though fixed code supplies scoring. MEM-RTE-brief-delivery supplies lexical ranking and forced identity matches; MEM-OBJ-query-log supplies a demand-priority view, not demonstrated adaptive compression. Standing, validity and mode fields also route which work or content is selected. Arbitrary policy JSON has known SQLite storage and a symbolic envelope, but its internal semantics and later host effect remain unresolved in MEM-OBJ-learning-policy.

Software admission without per-content human approval is automatic even when a person starts a survey. Explicit confirmation fields, actor labels and direct-user provenance do not authenticate manual control. RT-RTE-decision-admission and RT-RTE-policy-admission preserve that distinction. Initial ledger control remains unresolved rather than inherited from known subsequent writers.

Transformation and later delivery

Whole-body explanation replacement and claim successors support evolution; interval closure and withdrawal support invalidation. Persisted brief selection and role aggregation reduce retained material and support consolidation. Acceptance/activation promotes a draft or qualified candidate into current authority, without implying deployment or correctness. EPI-RTE-review-aggregation derives retained keys from judgments, not new source truth. No synthesis value follows from a model being asked to explain code; EPI-RTE-explanation-record leaves actual body entailment unresolved. Hash checks and primary-key uniqueness do not establish deduplication.

Requested locus/raw/brief lookup is pull. The automatic provision of stored task context to a design lens is the push operation inside the wider requested workflow: compilation selects retained candidates, planning narrows sections, and dispatch delivers the frozen packet. The signal units name actual selection inputs rather than treating any request ID as an identity-based selector. Forced registry matches, optional lexical matching and coarse lens/item filtering remain distinct. MEM-RTE-brief-delivery records both this chain and its forced-before-mode-filter caveat. RT-RTE-locus-read instead fulfills a retained-content request; its byte budgeting alone does not make that route push.

Generated views add file-backed later consumption; they do not replace durable SQLite/prose authority. RT-RTE-docs-publication retains normal, clean, verification-disabled and standalone paths. Projection checking does not certify codebase knowledge, and a successful clean promotion need not include readback when verification is disabled.

Learning uncertainty and inventory coverage

Trace-derived lineage is not equivalent to trace learning. MEM-OBJ-learning-lessons records automatic terminal-artifact extraction beside separately authored propositions/configuration. RT-RTE-policy-admission establishes qualification, durable active representation and internal reader comparison. MEM-OBJ-learning-policy leaves future host interpretation unresolved; the active serializer omits the proposition and evidence ledger. These facts do not close a qualifying later behavior-shaping route, but neither establish its absence.

Usage traces and adversarial/review history are likewise not automatically qualifying learned artifacts. MEM-OBJ-access-log and MEM-OBJ-query-log establish priority inputs; WF-RTE-adversarial-review establishes retained criticism and prescribed packaging use, not automatic induction from conversational traces. No accepted evidenced-absence record licenses a local trace-learning no. Consequently trace learning and trace source remain unresolved with no union value, rather than negative or inapplicable. Session identifiers alone cannot distinguish conversational logs, tool traces, event streams and linked state/action/outcome trajectories.

The unresolved inventory includes the additional registered maintenance, evaluation, adapter and driver chains retained in the analysts' limitations, not merely active policy. Their missing payload, control and downstream-consumer facts prevent complete-value conclusions even where a strong local witness exists. The policy and dispatcher records anchor these limits; inspected disposition, contradiction, diagnosticity and sealed-review mechanisms are included through the epistemic records, not discarded because memory's earlier inspection was narrower. No overall absence, improvement or complete enumeration is inferred.

Source-read log

None. This classification used only the supplied accepted members and reconciliation; no named record ambiguity required a frozen-source read.