MDCP memory profile

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

Comparison rationale

This comparison classifies accepted records, not a new source inspection. All ten axes remain partial because the memory account includes uninspected complementary products and evaluator-helper artifacts. A whole-system functional boundary does not establish complete classification coverage. Local known units describe the named mechanisms, not an assertion that every possible controlled value has been enumerated.

Documents, derivatives and policy

MEM-OBJ-shards separates mutable project contracts from installed vendor policy. Text has instruction or knowledge authority at interpreting readers; links and ordering have symbolic routing effects at the compiler. MEM-OBJ-research further separates observations from positions and normative clauses. The observation supplies evidence, not a command. Research-record and ADR supersession support history-preserving invalidation; deletion of ordinary current prose alone would not.

MEM-OBJ-compiled, MEM-OBJ-refs and RT-RTE-compile-check establish file-based derivatives and concrete later CLI/site consumers. A refs title is not a semantic memory payload, and its heading/path fields have fixed consumers. EPI-RTE-code-evidence adds computed lexical locations; EPI-RTE-inline-inserts adds imported bodies. Neither is independent warrant for surrounding prose. EPI-OBJ-context-size and EPI-RTE-context-size add retained measurement fields and their deterministic production route, not observed agent savings.

Static shipped configuration is not itself accumulated memory. The authority inventory includes configuration/checking guidance only on the recurrence-update path described in RT-RTE-doc-review, and includes indexes/refs as changing access structures. Likewise, the host model is not assigned model-weights storage or parametric memory simply because inference consumes these documents. The compiler's process-local shard cache is excluded from later-invocation memory; retained optimizer state crossing separate inference invocations is included.

Admission and maintenance

MEM-RTE-shard-authoring and RT-RTE-feature-work prescribe agent admission without mandatory per-content human approval. That supports claimed automatic agency for the agent-mediated route, not manual agency from user initiation. Conditional approval, unknown initial content control and the further author action installing a returned optimizer description remain unresolved admissions. Mechanical import, compilation, peer fixing and internal optimizer selection independently support wired automatic agency.

Review merge/dedup operations remain afforded, and current-state authoring and history-preserving supersession remain claimed. The optimizer's candidate revision is wired evolution. No synthesis is inferred from generated wording or research positions without a recorded input/output relation establishing a new claim; no decay or promotion is inferred from deletion, backups or rebuilding. RT-RTE-hook-extension and MEM-RTE-peer-fix establish mutations but not every opaque transformation's semantic operation.

Reconciliation supersedes the epistemic peer-fix declaration with MEM-RTE-peer-fix. This profile uses that canonical route. The amendment changes citation identity, not a value or basis. Compile body and persisted compiled files remain distinct lifecycle scopes; candidate descriptions remain separate from installed vendor skill policy. No other supersession changes a comparison value.

Read-back and trace-fed policy updates

The host asks for shards, and caller-invoked compile/refs/site operations ask for retained files. Their requested delivery is pull even where deterministic selection occurs inside the chain. File paths and manifest identity matches therefore do not establish identifier-based push. The parent skill's static session-start installation is not accumulated-memory injection.

RT-RTE-skill-evaluation and EPI-OBJ-skill-description establish a different route: the loop supplies retained candidate and failure/history state to subsequent inference invocations automatically. Fixed failure/status and history-field filtering support coarse push. The scores separately have ranking authority at candidate selection. Neither scalar scoring nor a best-description label supports embedding, lexical relevance or model-judgment push selection.

Execution feedback feeds automatic policy-description revision, which later proposal/evaluation invocations consume. That supports wired trace learning within the bounded maintainer loop even though final production installation is separate and operation/benefit are unobserved. The feedback has learning authority at the update consumer; the resulting candidate has instruction authority at its next trigger consumer. Local retained files and cross-invocation loop state supply the relevant horizon; there is no inference of model-weight updates or permanent production modification.

The subprocess execution-feedback part supports tool-traces as an original source. Static fixture labels do not become session logs by entering a model call; runtime occurrence alone does not justify an additional event-streams classification. Optional logs and broader grading inputs retain their unresolved provenance. Review recurrence supports a claimed guidance-update route but does not identify original trace content precisely enough to classify its trace source. No evidenced-absence record licenses a local or system-wide trace-learning no, and unresolved alternatives cannot be made inapplicable by a known route.

Evidence and remaining coverage

Each finding keeps its own claimed, afforded or wired basis. No observed or causal basis is supplied by this profile. Unknown complementary products, UX/tutorial details, evaluator artifacts, callback semantics, initial/final admission control and precise backup/publication maintenance prevent stronger inventory conclusions. External host compaction and provider state are boundary exclusions, not negative findings about MDCP.

Source-read log: none. The supplied accepted members and reconciliation sufficed; no source ambiguity required reopening a cited source file.