Context Constitution memory profile

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

Comparison rationale

All positive findings have claimed basis. The accepted records describe a policy and external host affordances, not inspected memory implementation or execution. A unit can therefore have known classification at a declared path without wired or observed evidence. Partial axes retain opaque initial content, provider storage, mixed skill assets, alternate write gates and delegated consumers rather than excluding them to obtain completeness. The inventory and those limits come from the memory objects and annotations on RT-RTE-memory-admission, RT-RTE-context-disclosure, RT-RTE-history-compaction and RT-RTE-subagent-context.

The summary identity supersession affects every summary-related axis: citations use canonical MEM-OBJ-summary-message, not the superseded epistemic summary declaration. The amendment changes identity resolution, not evidence strength or fidelity. No other supersession changes a value. EPI-OBJ-durable-learnings remains a content class within the update route, not a replacement for the block, external-memory or projection objects.

Logical memory blocks map to service objects, while their working projection maps separately to files and a repository. A named message database does not establish relational storage, SQLite or vectors. Summary buffer placement does not establish durable backend storage or a history of summary versions. These limits are retained in MEM-OBJ-memory-blocks, MEM-OBJ-mem-fs, MEM-OBJ-message-history and MEM-OBJ-summary-message.

Automatic baseline supply is coarse push; requested full-content delivery is pull. The foreground relevance judgment in RT-RTE-context-disclosure selects what to request, not an additional unsolicited delivery. References and paths therefore do not establish identifier-based push. Curated discovery material has routing force at that decision, while retrieved content supplies advice or context; standing system memory separately supplies instructions. No ranking, validation or enforcement value follows from a catalog, relevance judgment or file-propagation gate alone. Opaque skill consumers remain unresolved under MEM-OBJ-skill-assets.

Compaction qualifies as consolidation and trace learning because an automatic trace-fed summary becomes context for later inference. This mapping does not assert semantic preservation: the epistemic account's indeterminate content relation and missing fidelity checks remain compatible with a declared summarization operation in RT-RTE-history-compaction. Its mixed session/tool origins come from MEM-OBJ-message-history. Retained traces have learning authority at that update consumer, separately from the summary's knowledge authority at later inference. Neither authority nor the trace-learning value establishes improved capacity.

Foreground text revision, compression and deletion support claimed evolve, consolidate and decay operations on the described maintenance branch. Merely moving material out of the active window is not forgetting it, and pruning confidence is not semantic rejection or history-preserving invalidation. The unresolved transformations of Reflection and Defragmentation receive no operation by role name. The relevant boundaries are RT-RTE-memory-admission, EPI-RTE-pruning-assessment and RT-RTE-subagent-context.

Source-read log: none. Classification used only the supplied accepted members and their reconciliation; no frozen source file, incumbent profile or reference profile was read.