Compound Engineering Plugin memory profile

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

Comparison rationale

The comparison covers accumulated operational state as well as project guidance. Private controller receipts matter because later recovery consumes them. Unchanged skill installation and resource registration are not themselves memory writes. Revised methodology is included only through the repository authoring pathway, where a changed retained method can guide a later author or host invocation. The principal inventory and its unresolved alternatives follow the memory report's coverage account, not a claim that the whole-system boundary was exhaustively audited.

The five reconciliation supersessions affect solutions, pack rules, glossary, pack resolution and planning delivery. Their canonical citations here are MEM-OBJ-solutions, MEM-OBJ-pack-rules, MEM-OBJ-concepts, MEM-RTE-pack-resolution and MEM-RTE-learning-retrieval. Capture grounding and refresh remain distinct parts of RT-RTE-learning-admission, not duplicate routes. No unresolved reconciliation conflict changes a value.

Evidence strength remains attached to the mechanism. The wired pack-cache function establishes file/Git acquisition, not rule quality or observed reuse. Wired controller validation establishes vetoes on the bundled transaction route, not a universal sandbox. The wired session extractor reduces archive material; the historian-to-learning-to-later-reader chain remains claimed. The mechanical script in EPI-RTE-capture-grounding cannot upgrade semantic admission or retained prose to wired truth checking. Full grounding is postwrite and adjudicated; Lightweight lacks that semantic pass.

Pack prescriptions have knowledge authority at their planning/review consumers, because they are explicitly treated as evidence rather than workflow instructions. Their metadata and relevance influence ranking even without a learned ranking policy. Controller fields instead supply validation, enforced veto and routing at fixed-rule consumers. Experiment experience supplies learning authority at the digest update; the digest supplies knowledge at later hypothesis generation. These paths do not demonstrate improved capacity.

Starting a workflow is separate from approving its retained content. Silent glossary edits, automatic session enrichment, model-written handoffs and batch digests are automatic admission mechanisms. Conditional refresh approvals and consent to discoverability instruction edits are manual mechanisms. Destination selection alone does not convert a pack or handoff content write to manual. Initial store/config control and methodology merge/admission remain unresolved.

Consolidation is supported by selection/compression of archived sessions, continuity and experimental experience, and by merging retained guidance. Faithfulness remains uncertain without turning every model summary into synthesis. No input/output instance establishes that a generalized pattern creates a claim absent from its inputs. Evidence-supported deletion withdraws active reliance and forgets live corpus content while Git retains history. Stale metadata does not establish automatic suppression at readers; age alone is not decay policy.

Read-back is operation-specific. The researcher receives a request for retained evidence and returns selected findings; this is pull, not a second push because software or a model performs selection. The same distinction holds for requested pack roots, history enrichment, explicit handoff selection and next-batch digest reads. Identifiers and lexical relevance on these paths do not support push-signal values. Host injection and changed-method loading remain unresolved, so the system is not classified as pull-only and the signal axis is not globally inapplicable.

Trace learning uses the profile's write-route meaning. MEM-RTE-session-capture, MEM-RTE-handoff-resume and MEM-RTE-experiment-digest each describe automatic experience-to-retained-artifact-to-later-consumer chains. A continuation snapshot qualifies despite its temporary storage and lack of generalization. Optimization histories link hypotheses/changes and measured outcomes, supporting trajectories alongside tool feedback; a timestamp or serialization wrapper does not add event-streams. Host-compacted summaries retain uncertain original provenance. No bounded absence record supports trace-learning no, and the authoring/evaluation pathway alone does not establish an automatic live-skill update.

Unresolved retune, dogfood, ideation, live-polish, feedback and sweep state, additional glossary/solution consumers, full recovery internals and host memory writers remain included limitations. Their missing leaf semantics, admission ownership, selection chains and original inputs prevent stronger coverage assessments. Display summaries and available search matches do not classify opaque consumed payloads. There are no observed-run or causally supported findings in this profile.

Source-read log: none. Classification used only the supplied accepted members, reconciliation, types and record contract. No source files, incumbent profiles or reference profiles were read.