CocoIndex memory profile

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

Comparison rationale

All findings classify the supplied records, not fresh source evidence. No source files were read for this mapping. No supersession changes a value: reconciliation retains the component memo, function memo, user-state and tracking objects as distinct referents.

Storage and encoding. LMDB maps to kv; PostgreSQL maps to rdbms, including when it stores vectors. Dense embedding parts are parametric, separately from natural-language text and symbolic lineage. The alternate Qdrant consumer supports a vector-store finding without establishing the whole image producer. Opaque serialized returns are not classified from their diagnostic summaries. See MEM-OBJ-function-memoization, MEM-OBJ-doc-vector and MEM-RTE-alternate-search.

Local derivations and control. Imported content and its local compilation are separate lineage paths. Initial state control and external embedding/summary provenance remain unresolved. Operator initiation does not turn software admission into manual approval. Known existing-entry replacement supports evolve; physical deletion, index construction and canonical-map names alone do not establish the other curation operations. See MEM-RTE-state-access, RT-RTE-live-maintenance and EPI-RTE-pair-disposition.

Consumers determine force and direction. Dense vectors and filing features rank candidates; selected textual results supply knowledge rather than standing instructions. Internal envelopes route work and supply cache-validity criteria. Explicit retrieval is pull even though code performs selection. Automatic substitution of retained results for a computation request is push: identity matching actually selects the delivered memo, rather than merely naming an entry in a display. Uninspected consumers prevent treating all remaining directions or signals as absent or inapplicable. See RT-RTE-query-index, MEM-RTE-function-reuse and MEM-RTE-inspect-state.

Trace-fed context is not demonstrated improvement. The session parser automatically derives retained textual records and dense retrieval representations from conversational traces, and a later invocation selects and displays those derived snippets. That chain supports wired trace learning under this profile's artifact-and-consumer definition, without inferring an answering agent, improved capacity, factual endorsement or newly generated context summaries. Its source is session logs, not event streams merely because checkpoints have identities. Uninspected upstream context generation and opaque alternatives remain local unknowns. See MEM-OBJ-session-embedding and MEM-RTE-session-search.

Reconciliation limit. The unresolved attribution of pair-decision retention to a component memo versus a decorated function entry prevents choosing that decision's specific internal container. The profile preserves the independently recorded symbolic decision, inference derivation and canonical-selection consumer without resolving the conflict or borrowing a stronger status. Generic LMDB memo-storage findings concern the separately established memo objects, not a repaired pair-decision dependency. See EPI-RTE-pair-proposal, RT-OBJ-component-memoization and MEM-OBJ-function-memoization.

All ten axes remain partial. Positives establish local existence, not complete enumeration across opaque caller content, alternate producers and uninspected downstream consumers. There is no observed-run or causal evidence in these classifications. The transient contributor results and current-call corrective messages do not become retained instruction or learning-authority findings merely because they can influence an immediate next step. See EPI-RTE-hook-result and EPI-RTE-pair-check.