Voyager memory profile

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

Comparison rationale

This profile classifies accepted records, not a new source inspection. All positive findings have a wired implementation basis. No supplied execution or causal evidence upgrades them to observed operation or demonstrated benefit. The inventory distinguishes accumulated objects from current parser output and static helper code: MEM-OBJ-skills-code, RT-OBJ-program-code and RT-OBJ-control-primitives.

Reconciliation supersedes the epistemic chest and description objects and the epistemic QA, chest and progress routes with their memory counterparts. The classifications therefore cite MEM-OBJ-chest-memory, MEM-OBJ-skill-description, MEM-RTE-qa-cache, MEM-RTE-chest-update and MEM-RTE-progress-update. The combined epistemic QA container is not substituted for its separate answer-map and question-index parts. The distinct description-generation subroute remains EPI-RTE-skill-description.

The executable library is standing symbolic direction at Node, classified as instruction rather than enforcement. A definition can guide execution when called without vetoing arbitrary code. The model-facing copy instead supplies knowledge/examples; the description index supplies ranking. These are separate forces and consumers, not a claim that the executor actually called every delivered function. See RT-BAP-skill-execution and MEM-RTE-skill-retrieval. Unknown additional effects inside unsampled bodies keep authority partial.

Requested semantic retrieval remains pull even though the driver later formats its result. The separate automatic whole-library channel is push. Likewise requested QA lookup differs from automatic curriculum display of eligible pairs. Identifier lookup and embedding similarity inside those requested returns do not establish identifier-based or embedding-based push. Fixed display eligibility, budgets and whole-map supply establish coarse push selection; they are not a new relevance inference. See MEM-RTE-skill-retrieval, MEM-RTE-qa-cache and MEM-RTE-context-refresh.

Manual task approval is local to the manual critic gate on the retained program. It does not turn later generated descriptions, chest observations or progress-list mutation into manual content decisions. Conversely, an arbitrary caller selecting a checkpoint does not establish automatic or manual original admission for each entry. These unknown alternatives remain in write agency and lineage rather than inheriting the live automatic route: RT-RTE-task-judgment, RT-RTE-skill-admission.

Curation classifies implemented changes to retained content. Same-name skill replacement evolves an active artifact; progress cleanup deduplicates and withdraws failure membership while recorder history remains. Replay consolidates retained event information into reduced accounting. New-candidate description generation, embedding construction and ordinary cache acquisition do not alone establish curation; semantic indeterminacy does not establish synthesis. See MEM-RTE-progress-update, MEM-RTE-event-replay and EPI-RTE-skill-description.

Trace learning is an automatic durable-write classification, not proof of capacity gain or model-weight training. Skill, chest and outcome-history chains support local yes findings. Event archives have a later limit-control consumer, but the records do not establish a separately durable derived learning output; block sets and replacement action frames likewise have insufficient durability warrant for a stronger classification here. QA observations may shape questions while answers come from model knowledge; its exact trace-fed contribution remains unresolved. No local no finding or system-wide negative is inferred from these gaps. See RT-RTE-skill-admission, MEM-RTE-chest-update, MEM-RTE-progress-update, MEM-RTE-event-replay, MEM-OBJ-block-records and MEM-RTE-qa-cache.

Source-read log: none. Only the supplied members and governing contracts were read; no frozen target files were reopened. Historical trial payloads, embedding identities and initial content-control facts remain missing recorded evidence, not absent mechanisms. The known storage and delivery architecture does not make those content/provenance axes complete.