Apache Maka memory profile

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

Comparison rationale

Accepted records and scope

This profile classifies the accepted record inventory, not the whole runtime's storage or capability surface. Each axis is partial because its named inventory retains unresolved included paths. Known local units do not establish complete whole-system coverage. Implementation is wired, not observed use or benefit. The retained-state inventory and its gaps are supplied by MEM-RTE-local-memory, MEM-RTE-index-maintenance, RT-RTE-executor-plugin and RT-RTE-plugin-admission.

The reconciliation supersedes EPI-RTE-local-memory-use with MEM-RTE-local-memory; all corresponding route classifications cite that canonical record. It leaves bundle/item containers alongside their more precise parts. No other supersession changes these values.

Parts, provenance and admission

MEM-OBJ-text-checkpoint and MEM-OBJ-provider-checkpoint are alternative outputs of RT-RTE-context-projection. The text branch establishes summary form and consolidation; it cannot classify the hidden provider payload. Neither checkpoint's writer seam establishes its physical storage. A visible wrapper can support symbolic compatibility routing while leaving payload form and downstream force unresolved.

MEM-OBJ-bundle-metadata records manual/editor surfaces and user-labeled approval, but not a human decision at every caller. The write-agency unit therefore remains not-determinable, without deleting the independently automatic fact, checkpoint, index and Matter admissions. EPI-RTE-memory-canonicalize supplies a fallible evidence-consuming support criterion; it is not an external truth oracle. Consolidation follows the implemented extraction target, not a claim that every output faithfully preserves every source assertion. Arbitrary index writing supports authored lineage and evolution; it does not alone prove synthesis of a proposition absent from the inputs.

Consumers, direction and trace-fed updates

MEM-RTE-recall-history is request-driven even though its code ranks literal matches. Its lexical selection does not establish an inferred-lexical push signal. In contrast, MEM-RTE-local-memory automatically supplies active/session-matched entries at composition: coarse filtering and actual session-identity matching both matter. A requested Read address and a Matter wake's file paths do not by themselves establish identifier-based body push. Saved review reuse has its own software consumer, distinct from later reading of the feedback narrative in MEM-RTE-matter-state.

Learning authority is attached to recorded experience where it feeds durable updates; knowledge or instruction describes the downstream retained result. MEM-RTE-fact-extraction and textual RT-RTE-context-projection establish qualifying automatic trace-fed writes and later model consumption without establishing improved capacity. The narrower EPI-RTE-goal-selection likewise retains conversation-derived next-turn steering; it closes part of the memory report's goal inspection gap, not all Goal/Todo/WorkHub coverage.

MEM-RTE-history-extract saves files and returns a preview without establishing automatic full-body adoption. MEM-OBJ-review-archive establishes a later client query, not autonomous behavior revision. WF-RTE-patch-publication establishes retained patch publication and later listing, not parent application. These limits prevent inferring qualifying trace learning for every generated product. No absence record supports a global no or an absent curation value.

Original content, not serialization, determines trace source. User messages used by fact canonicalization remain session logs; runtime-event wrappers are not independent environmental event streams. Text compaction also consumes recorded tool results. Index and Matter mixed originals retain source uncertainty rather than acquiring an event-stream or trajectory label from their subsystem names. See EPI-RTE-memory-evidence, RT-RTE-context-projection and MEM-RTE-index-maintenance.

Source reads

No source files were read for this profile. All classifications and limitations use the supplied accepted records and their retained evidence passages. No new records, quotations or evidence are introduced.