MGI-Mind memory profile

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

Comparison rationale

All findings classify accepted records, not fresh source evidence. No source reads were performed. Bases describe inspected wiring or the separately afforded host-file consumer; no observation or causal benefit is asserted.

Reconciliation supersedes the duplicate epistemic memory, fact, procedure, skill, procstats and external-signal objects with their memory-member declarations. Citations above use those canonical declarations. The provenance part and session-summary part remain distinct from their containers. Route overlaps retain their separate functions rather than being treated as stronger evidence of execution.

Qdrant payload persistence maps to service-object; its operative retrieval indexes additionally map to vector. This does not infer Qdrant's internal disk technology. Dense vectors are parametric representations, while sparse hashed lexical indexes are symbolic. Local model artifacts are not use-updated memory. Arbitrary vault values remain unresolved despite known file storage.

The ordinary wrapper delivers requested note text as knowledge, omitting hit identifiers and scores; see WF-RTE-response-delivery. Procedure and skill bodies instead supply guidance. Derived outcomes exert ranking authority at code consumers, while experience has learning authority at the update consumer. These are separate paths, not proof of reliable learning or changed answers. Pinned placement alone does not classify arbitrary block text as instruction.

Read-back separates explicit host requests from background retest supply. Context selection and match inside requested operations are not independent pushes. Background bounded candidate selection supports coarse supply to a server evaluator; detailed priority remains unresolved. The known pull paths cannot make the entire signal axis inapplicable.

Curation distinguishes duplicate removal, coldness decay, promotion, withdrawal and replacement from standing-fact consolidation. The conditional worker extracts triples from retained memory text and submits them to durable fact admission; harvested session messages can feed that path. This supports wired consolidation without observed outputs or semantic fidelity. The structural-check limits in EPI-RTE-extraction-parser constrain semantic warrant, not the implemented curation classification. No faithfulness defect is redescribed as synthesis.

Outcome counters, access-derived maintenance, query-derived drift state and conditional transcript-to-fact extraction establish trace-fed durable behavior-shaping routes. MEM-RTE-transcript-harvest supplies user/assistant session-log input; RT-RTE-auto-extraction supplies the feature- and installation-dependent automatic write; MEM-RTE-fact-lifecycle and RT-RTE-context-briefing supply later consumption. This wired route requires admission, not observed triples, entailment or benefit. Tool/thinking blocks are excluded, so this branch does not establish tool-trace origin. Reported success types do not authenticate tool execution or identify the original trace source. No evidenced-absence record supports a local trace-learning no. Session summaries are retained, but their production and control are not established by the session-end storage call alone. Initial and maintenance alternatives stay unresolved rather than disappearing from the inventory.

The fact typed-signal namespace gap remains local to that bridge; it does not erase playbook ranking or scalar drift state. Briefing checks an upper temporal bound, not complete stored intervals. Full restore, migration/backfill, caches and evaluator-result consumers are incompletely traced; these limits prevent complete-value claims, not the supported findings above.