Smriti MCP memory profile

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

Comparison rationale

The memory boundary is the use-updated note, event and access-structure inventory in MEM-OBJ-memory-body, MEM-OBJ-memory-meta, MEM-OBJ-trace-events, MEM-OBJ-markdown-index and MEM-OBJ-machine-index. Fixed package rules and RT-OBJ-server-instructions are not accumulated memory. In particular, their tool schemas do not supply a validation-authority finding for retained content.

The supplied reconciliation resolves the duplicate epistemic body, metadata, trace and access-reinforcement declarations to the corresponding memory records. The profile cites those canonical memory anchors. The runtime containers remain valid alongside distinct epistemic checking, selection, retention, maintenance and lifecycle parts; none of those parts is treated as a second retained payload. Supersession changes citation identity, not evidence strength.

Files are the persistent substrate. Adjacency mappings do not add graph storage, and lexical counts do not add vector storage. Fresh Python access dictionaries receive a separate in-memory finding because they are an access structure derived from retained notes, not merely executor state. Saved YAML has storage and symbolic-form findings but no read-back or behavioral authority: MEM-OBJ-machine-index establishes export rather than reload. The Markdown index has a different, explicitly requested consumer.

Natural-language context and fixed-rule lexical consumption are separate paths through the same body. EPI-RTE-recall-selection supports an afforded reference/advice role at the external reader, not observed instruction compliance. Retained fields and terms have wired ranking force even though the scoring policy itself is fixed code. Default archived/superseded exclusion is bounded enforcement, not authentication or a universal read ban; EPI-RTE-lifecycle-disposition preserves full-get and override routes. Trace material has an afforded learning role at the review-to-update consumer, distinct from its reference role and from automatic trace-learning qualification.

Local deterministic derivation remains known despite unknown initial file provenance. The missing-description doctrine does not erase the compiled-prefix path: EPI-RTE-maintenance-rewrite records its implementation and loss-of-condition counterexample. A lossy reduction remains consolidation rather than synthesis merely because it is unfaithful. The machine index's rebuild alone establishes no curation operation. Likewise, the name consolidate_memory does not resolve the external caller's semantic transformation; RT-RTE-trace-consolidation records supplied prose, not an owned summarizer, and create/update provenance differs from append.

Write agency is local to admission. The automatic branch decision, repair edits and MEM-RTE-access-reinforcement are code-controlled even after a caller starts them. Generic prose suppliers and external file editors remain unresolved rather than manual by default. Their unknown control prevents a complete write-agency union.

Requested retrieval remains pull even when code ranks, filters, expands neighbors and formats the response. RT-RTE-context-recall, RT-RTE-trace-consolidation and the two index records cover the delivery inventory. The read-back signal axis is therefore inapplicable within this shipped boundary. Full retrieval remains available when compact recall omits body content, source references or reasons; no complete-content or actual host-use conclusion is implied.

The trace axes retain a narrower uncertainty than the behavioral-authority axis. An experience-to-prose affordance does not resolve whether a human or model admitted a particular semantic update, and unverified trace identifiers do not establish its input lineage. Automatic access reinforcement is independently wired, but its recorded input is a selected ID and prior metadata, not an established recorded trace derivation. Neither the absence of raw-trace writing on that route nor lack of observed improvement supplies a system-wide trace-learning negative. The accepted records consequently support neither a complete positive classification nor bounded absence, and trace source cannot be declared inapplicable.

Source-read log: none. This classification uses only the supplied accepted records and reconciliation; no frozen-source ambiguity required an additional read. No finding has observed or causally supported basis because the accepted set supplies neither execution traces nor comparative experiments.