Trajectory-informed memory generation memory profile

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

Comparison rationale

All positive findings are claimed. The accepted records describe doctrine and attributed aggregate results, not inspected wiring, observed extraction/retrieval traces or a registered causal experiment. A resolved classification within a declared unit therefore does not assert implementation or operation. In particular, the form inventory is known at the documented retained-output boundary even though physical persistence and other admission/access mechanisms remain unresolved.

Reconciliation supersedes the epistemic trajectory, tip and metadata duplicates with the memory objects. This profile uses MEM-OBJ-trajectories, MEM-OBJ-tips and MEM-OBJ-metadata. Those amendments change citation identity, not classification strength. The description-index container remains valid, but separate MEM-OBJ-descriptions and MEM-OBJ-embeddings preserve the different text and numerical consumers. No supersession changes a positive value.

The storage union contains vector, not model-weights: dense access representations are retained outputs, while provider weights are producing/consuming components with no established update in this memory boundary. Likewise, parametric form refers to those dense representations, not evidence of parameter training. The unresolved relationship between content/purpose vectors and description indices prevents a complete schema claim, but does not change their shared numerical form. See MEM-OBJ-embeddings.

Instruction and knowledge distinguish prescription from explanatory context at the guidelines consumer. Learning instead concerns retained experience feeding a successor update at consolidation. Ranking concerns retained descriptions, vectors and outcome/priority attributes at actual selectors. Routing concerns metadata selecting retrieval scope. None grants API permission or demonstrates that an instruction changed an action. These paths are supported by RT-BAP-guidelines-channel, RT-RTE-tip-consolidation, RT-RTE-llm-selection and EPI-RTE-conflict-check.

The curation union intentionally does not equate the source's word synthesis with the controlled value synthesize. RT-RTE-tip-consolidation establishes requested semantic merging; EPI-RTE-conflict-disposition retains indeterminate successor semantics. Combining complementary steps does not establish a claim absent from the inputs. Conversely, the extraction example's broader conjecture is established in EPI-RTE-attribution-generation, but that production from a trace is not by itself an operation over already retained memory. No positive synthesis finding is therefore established for curation. Expiry, decay, reversible invalidation and promotion cannot be inferred from timestamps or source-outcome precedence.

Both retrieval alternatives supply selected tips automatically on a task trigger. Executing an internal query is not a foreground agent's request for memory. Embedding selection and model-derived query criteria remain separate signals; top-k settings bound item counts only, and retained identifiers alone do not establish identity-targeted delivery. See RT-RTE-cosine-retrieval and RT-RTE-llm-selection.

The trace-learning positive classifies automatic production of durable cross-task guidance, not improved capacity or exercised self-improvement. The original input contains conversational material and action/result feedback organized as complete trajectories, so all three source findings survive the union. Optional labels and self-inferred outcomes alter warrant, not that original trace origin. Neither aggregate gains nor counterexamples upgrade these design-level findings; RT-CLM-reported-results explicitly preserves configuration-specific harm and missing component attribution.

Source-read log: none. Classification used only the handed accepted records and reconciliation; no frozen-source ambiguity required a source read. No records or evidence were added.