AgeMem memory profile

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

Comparison rationale

This profile classifies the accepted documentary records. Every positive has claimed basis: pseudocode, formulas, prompts and attributed experiments do not establish executable wiring, inspectable operation or causal support. Coverage assessment is separate. A named mechanism can be completely classified at design level while its execution remains uninspected.

The three identity supersessions concern the content object, store-maintenance route and explicit retrieval route. Their canonical citations here are MEM-OBJ-content, MEM-RTE-store-maintenance and MEM-RTE-requested-retrieval. They change citation identity, not evidence strength or acceptance. The active context container remains distinct from its summary part; the embedding remains distinct from the encoder; experience remains distinct from updated weights. No unresolved inter-report conflict or value-changing supersession was supplied.

An external store and an active message list do not identify a storage backend. The only positive substrate classification is model weights, at the update's retained result rather than its unspecified checkpoint files. Dense embeddings establish parametric form, not vector storage. These distinctions keep unknown physical persistence beside the relevant objects (MEM-OBJ-embedding, MEM-OBJ-summary, MEM-OBJ-experience, RT-CMP-agent-policy).

The knowledge path is retained prose to later policy context. The learning path is experience to optimizer, and the downstream parameter path selects later actions. Ranking belongs to the retained embeddings or message text consumed by similarity selectors; routing belongs to search/target metadata and the policy's operation selection. Static prompt instructions and schema validation requests are outside accumulated memory. None of these paths establishes epistemic endorsement or human approval (MEM-BAP-retained-context, MEM-OBJ-access-metadata, RT-RTE-policy-update).

Summary's declared compression supports consolidation even though output fidelity is not demonstrated. Arbitrary generated Add/Update strings have no established semantic relationship to their inputs; their indeterminacy is not permission to assert synthesis or deduplication. Delete, Filter and reset forget material, but no history-retaining withdrawal route establishes invalidation. The classification therefore preserves positive transformations without turning maintenance rewards into evidence of correct repair (RT-RTE-context-summary, RT-RTE-context-filter, EPI-RTE-language-content, MEM-RTE-store-maintenance).

The explicit retrieval return is pull. Stage 1 pre-action retrieval is push, with sampled message, cosine selector and selected stored entries named together. Context carryover and training consumption are separate push operations; they are not additional pushes inside a requested retrieval return. Identifier-based mutation targeting does not establish identifier-based delivery. Unknown whole-store serialization and deployment loading remain unresolved units, so a known pull route cannot make the signal axis inapplicable (MEM-RTE-requested-retrieval, RT-RTE-stage-retrieval, RT-RTE-task-rollout).

Trace learning here means an automatic trace-fed write with a retained behavior-shaping result and later consumer, not demonstrated capacity gain or exercised self-improvement. Summary meets this criterion within an episode; conversation-fed storage and GRPO supply other positive paths. The summary write's trace-source coverage is partial: session-log and tool-trace positives do not resolve the original provenance of supporting information, synthetic distractors, retrieved content or earlier summaries included in selected context (RT-RTE-task-rollout, EPI-OBJ-context-content). Tool requests recorded as actions support tool-trace origin, while linked states/actions/outcomes support trajectory origin. The records do not establish complete raw tool-result logging, static-input provenance or initial backbone training provenance. No bounded absence record licenses a system-wide negative (MEM-OBJ-experience, RT-RTE-context-summary, RT-RTE-policy-update, MEM-OBJ-content).

Source-read log: none. This classification uses only the supplied accepted records and their annotations; no ambiguity required opening the frozen capture.