MiroShark memory profile

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

Comparison rationale

This profile classifies accepted records, not source execution. Every positive finding is wired implementation evidence; there is no supplied observed run or causal experiment. All axes remain partial because local witnesses do not close the retained-memory inventory or its auxiliary and host alternatives.

The comparison separates graph text, symbolic selection metadata and dense access vectors. Belief scalars are symbolic state, not model parameters. Persona/configuration files and caches count through reuse, not because all files are memory. Current observation filtering alone is outside the retained boundary; its later accumulated/derived uses remain covered by the round and belief routes. See MEM-OBJ-agent-observation, MEM-OBJ-belief-state and RT-RTE-prepare-population.

Consolidation does not require a faithful observed output. The round and community mechanisms compress existing retained content, while entailment of particular generated claims remains unresolved. Neither an operation called synthesis nor missing fidelity evidence establishes a new claim absent from inputs. Entity resolution supports deduplication; exact-replay skipping in contradiction checking does not make relation creation deduplicated. See MEM-RTE-round-history, MEM-RTE-community-rebuild, EPI-RTE-entity-resolution and MEM-RTE-graph-ingestion.

Learning authority at the update consumer is separate from knowledge or instruction at the subsequent agent. Qualifying trace learning includes resident continuation summaries because later invocations receive the retained result; cross-run serialization and observed benefit are not prerequisites. Saved trajectories and report traces are not automatically training inputs. Approximate exposure limits what belief feedback says about actual perception, without erasing its implemented update. See MEM-RTE-belief-feedback, MEM-OBJ-cross-platform-log and MEM-OBJ-reasoning-trace.

Requested graph search remains pull even though code ranks the response. Round scheduling independently pushes history, stance and cross-platform digests. Pending MCP results instead fulfill model-authored tool requests: delayed delivery and agent-ID recipient matching do not establish another push or an identifier push signal. This requested-result unit is pull, with its push signal inapplicable; independent round-context push and identifier findings remain. Embedding and lexical retrieval establish ranking authority but not inferred push signals. Temporal withdrawal enforces only covered edge-search paths; node/fallback/historical paths and copied summaries remain separate. See MEM-RTE-graph-retrieval, EPI-RTE-graph-ranking and RT-RTE-agent-mcp.

Reconciliation resolves duplicate relation, belief and community objects to memory declarations and resolves duplicate round-compaction, belief-revision and community-generation routes to their memory counterparts. Those canonical destinations supply all affected findings above. The narrower round-prose object remains a valid contained part, not a duplicate of the whole round store. No recorded unresolved cross-report conflict requires choosing a stronger status. Report synthesis receives truncated returned sections; it is not a semantic acceptance gate, and parallel workers do not consume sibling drafts. See WF-RTE-report-delivery and RT-RTE-report-sections.

Source-read log: none. Classification used only the supplied accepted members and reconciliation; no source ambiguity required an additional read. Missing initial approval, exact remote payload provenance, complete analyzer/market-media processing and auxiliary consumer inventories prevent stronger assessments. These uncertainties do not establish absence, manual admission, universal withdrawal or improved future capacity.