Claude Context Guard memory profile

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

Comparison rationale

The profile classifies specified retention and later consumption, not successful operation. Model-operated writes, selection and maintenance have a claimed basis. Inspected shell copying has a wired basis. Exported helper consumption remains afforded: neither export nor a proposed two-way test establishes an actual call. No observed or causally supported finding is available. The inventory and its limits come from the accepted objects and route annotations, especially RT-RTE-save-session, RT-RTE-start-session, RT-RTE-itemise-code and RT-RTE-compact-backup.

Reconciliation supersedes ten epistemic object declarations with the corresponding memory objects. All profile object citations use those canonical memory declarations: comments, session history, handoff, tasks, decisions, lessons, QA, sidecars, helpers and audits. The distinct epistemic check/admission/inference/filter parts and runtime producer containers remain valid. Supersession does not upgrade an evidence basis or erase epistemic limits. In particular, EPI-RTE-feature-admission distinguishes human QA admission from ordinary model-written records, and EPI-RTE-lesson-inference leaves semantic production indeterminate.

Behavioral authority belongs to a consumed part and effect. Project records supply knowledge; decisions and directives supply instruction; experience has learning authority at an update consumer; clean-state/index/address metadata supplies routing or validation. A helper's possible enforcement is kept at its documented future caller. No shell gate is inferred from imperative prose. Relevance selection alone is not ranking. The helper and backup branches retain missing-consumer limits rather than borrowing the startup consumer of a different object. See MEM-OBJ-helpers, MEM-OBJ-resume-state, MEM-OBJ-sidecar-index and RT-RTE-compact-backup.

Requested startup and code-range reads are pull. Identifier matching in a requested grep chain therefore does not produce an identifier push finding. The claimed host auto-read of mutable project guidance is a separate coarse push; static shipped doctrine is outside the compared memory. No inspected consumed compaction payload is available. These distinctions follow RT-RTE-start-session, RT-RTE-itemise-code and RT-OBJ-claude.

Ordinary automatic continuation summaries qualify for claimed trace learning through their intended later consumer, even without semantic novelty or demonstrated improvement. Manual lesson installation is a separate admission path to the same LB object; it does not suppress the ordinary automatic positive. Conversational origin supports session-logs, but script versions and project observations alone cannot establish recorded tool-traces or linked trajectories. A copied ledger is not a new trace-learning artifact merely because its original production used conversation. See RT-RTE-save-session, MEM-OBJ-learned-behaviour and RT-RTE-harvest-lessons.

Instruction precedence remains unresolved for startup plan/audit loading, comment removal versus archival and rotation categories. Sidecar identifier parity does not establish description accuracy, and source restoration does not establish paired sidecar restoration. These limits prevent deterministic retention/read coverage or semantic-fidelity guarantees, not recognition of the specified file and selection paths. Missing initialized custom contents, original host-plan derivation, actual helper call sites and host-compaction state remain beside their units. All ten axes retain partial inventory coverage rather than excluding the opaque alternatives to obtain known. See MEM-OBJ-comments, MEM-OBJ-plans, MEM-OBJ-audits, MEM-OBJ-sidecar-index and RT-RTE-compact-backup.

Candidate novelty checks are not retained-memory deduplication. The task and lesson inventories supply admission references, but duplicate-candidate rejection does not establish merging or removing near-duplicate retained entries. Old task-row consolidation remains a separate claimed operation; it does not establish duplicate-merging semantics. See MEM-OBJ-task-registry, RT-RTE-harvest-lessons and RT-RTE-save-session.

Source-read log: none. Only the supplied accepted members, correction baseline and contracts were read; no source ambiguity required a frozen-source read. No records, quotations or new evidence are added.