llm-context-base memory profile

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

Comparison rationale

The profile classifies the accepted file loop, not all storage or guidance of its external host. The inventory is grounded in the object and route records, especially MEM-OBJ-knowledge, MEM-RTE-pattern-capture, and the annotated RT-RTE-session-start, RT-RTE-capture-file and RT-RTE-knowledge-query. Known coverage does not upgrade doctrine to wiring. Most findings remain claimed; documented integrations retain afforded bases; the converter alone supplies the relevant wired transformation/admission witness. No finding is observed or causally supported.

Reconciliation supersedes EPI-OBJ-sources with MEM-OBJ-sources, and EPI-RTE-pattern-capture with MEM-RTE-pattern-capture. Accordingly, this profile cites those canonical memory declarations. The narrower metadata transformation remains EPI-RTE-metadata-transform, alongside its valid invocation container RT-RTE-metadata-migrate. Other valid containers and their parts are not treated as replacement identities.

Retained parts and consumers

Filesystem and repository persistence can coexist. The repository value is conditional: MEM-OBJ-knowledge records ignored journal/private bodies and optional plain/encrypted Git recipes. Encryption is an encoding/access arrangement, not an additional controlled storage substrate. The optional sentinel and read-path log are local control/access state, not a readable representation of the host summary. The excluded host payload cannot be classified from the inspected sentinel. Opaque original-format sources and embedded artifacts remain included: their file substrate is known while their full forms, provenance and consumers are not.

Natural-language summaries and guidance remain natural language at the assistant consumer even when stored in YAML or Markdown. Fixed-rule metadata conversion and shell sentinel/path operations support separate symbolic consumption paths. MEM-OBJ-metadata and EPI-OBJ-metadata support this distinction. No external assistant weights are included as accumulated target memory.

Authority follows the actual effect. Outcome text supplies knowledge in decision deliberation; summaries/tags/status influence ranking; phase/index/date state supplies routing. The pattern update uses experience for learning, while its retained directions supply instruction at RT-BAP-instruction-load. These distinct paths do not prove improved capacity. Use-revised checklist/template guidance can supply validation criteria, but EPI-RTE-completion-check supplies only a procedural accuracy/completion criterion, not factual acceptance or executable enforcement. Initial static schemas and host permission settings do not become memory merely by being loaded.

Admission and transformations

Explicit human answers/edits/approvals are manual. Clear pattern capture and item-level model extraction are automatic despite upstream user initiation; ambiguous pattern admission, directory creation and optimization restructuring have explicit approval branches. MEM-RTE-pattern-capture and RT-RTE-context-review preserve those controls. The converter's per-file admission is automatic even though the operator launches the batch. Its rewrite is other-compiled lineage, not trace learning.

Consolidation describes selected observations and reduced source content. Optimization compaction concerns revised instructions, not opaque host conversation summaries. Retirement supports withdrawal of current reliance and reduced priority, not enforced exclusion. Draft-to-active completion supports promotion at the claimed lifecycle layer. MEM-RTE-lint-maintenance retains conflicting link/metadata/discard wording; the profile therefore asserts neither a universal schema validator nor a seven-day deletion mechanism. EPI-RTE-metadata-transform records byte/meaning loss, but fixed conversion or a faithfulness defect alone does not establish deduplication or synthesis. The semantically unspecified saved-answer route remains indeterminate rather than acquiring synthesize from its name.

Read-back and trace-fed writes

Requested answers are pull; startup/task guidance, overdue nudges and proactive precedents are push. The latter uses topic/domain and summary/tag signals plus model judgment; broad phase/age/sentinel conditions remain coarse. Named paths are not sufficient evidence for identifier-based targeted push. The file/line budget bounds delivered material, not storage or complete corpus coverage. Linked-code consultation and original reprocessing remain afforded consumers with unresolved request/selection detail, as recorded in MEM-BAP-linked-context and MEM-OBJ-sources.

Conversation-fed automatic retained facts/rules and derived knowledge qualify for trace_learning yes at the claimed layer because the records connect their session inputs, durable products and later consumers. This classification neither requires new claims nor asserts successful improvement. Manual branches and metadata conversion do not qualify; raw control/path logging alone also does not qualify. Unspecified outcome-writing and saved-synthesis/source-input chains prevent complete enumeration. No local no or absence is asserted without a bounded warrant. Original conversational input maps to session-logs, not event-streams because a selected observation is written into a training or operation log. MEM-RTE-compact-reload establishes only reload around an excluded opaque host compactor, not an additional trace-to-summary writer.

Source-read log

No frozen-source files were read for this profile. All classifications use the supplied accepted records and annotations, resolved through the supplied reconciliation. No evidence or record was added, and no report verification was reopened.