Memori memory profile

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

Comparison rationale

This profile classifies the supplied accepted records, not a newly inspected implementation. No source reads were needed. All implementation findings are wiring, not observed activation or causal benefit. Claimed hosted transformations remain claimed even beside wired client submission, local persistence and tool-result delivery.

Canonical identities and coverage

Reconciliation supersedes the epistemic fact, triple, summary, process-attribute and embedding object declarations with the corresponding memory objects. It also supersedes epistemic recall-selection, entity-withdrawal and feedback-delivery with their memory route declarations. This profile uses those canonical destinations. Their epistemic annotations and function-level assessments still constrain meaning; supersession establishes identity, not stronger evidence. In particular, MEM-OBJ-entity-fact retains mock/fallback emission and MEM-RTE-fact-selection retains selection rather than truth endorsement.

The broader EPI-OBJ-conversation-messages is not identical to persisted MEM-OBJ-conversation-message: transmitted host trace fields are not proven retained row fields. The benchmark artifact container and its answer/judgment function records likewise remain distinct. No superseded record is used as a classification destination.

Every axis remains partial because an opaque or uninspected included branch cannot be omitted to make the inventory complete. Graph/process storage is known without a separate later model-facing consumer. Feedback submission is known without retained feedback or an adaptation consumer. Benchmark import is known without the upstream memory-producing trace. These are different limitations, not evidence of absent mechanisms.

Part and consumer distinctions

The SQLite substrate also supports relational, graph and vector access structures; the values describe coexisting mechanisms rather than separate copies. Numerical fact embeddings are parametric representations, not learned model weights. Benchmark files are an independent retained route. The hosted service interface does not identify the private storage engine or representation. See MEM-OBJ-content-embedding, MEM-OBJ-knowledge-graph, MEM-OBJ-agent-memory and MEM-OBJ-benchmark-artifacts.

Facts can have ranking authority as lexical inputs and knowledge authority at provider delivery. A prior summary has learning authority at the update consumer and knowledge authority at the later answering consumer; its presence also routes extraction to the latest pair. These classifications neither certify truth nor establish improved capacity. Standing-order instruction is retained only as a claimed hosted output path, not inferred from a system-role envelope or static shipped skill. See MEM-OBJ-conversation-summary, MEM-RTE-fact-selection and RT-BAP-host-memory-guidance.

Transformations and read-back chains

Curation follows the recorded transformation. Summary overwrites are evolution; repeated summary removal is read-time deduplication; recurrence refresh changes salience through bounded candidate ordering. Destructive entity deletion is forgetting/decay, not history-preserving invalidation, and it leaves conversations replayable. Primary-key collision checks, triple rendering, embedding attachment and index rebuild alone add no curation value. Hosted extraction/summary consolidation remains claimed because its implemented semantics are unavailable. See MEM-OBJ-fact-access, MEM-OBJ-conversation-summary, MEM-RTE-entity-withdrawal and EPI-RTE-triple-rendering.

The SDK's backend request is pull; automatic fact/history selection into the provider call is push within that same invocation chain. This does not assert unsolicited background messaging. Intentional host tool returns remain pull even when the client automatically serializes them. Benchmark reload and automatic answer-context selection are similarly separate operations. Identifier findings on the SDK routes name actual entity/conversation/fact matches, not IDs merely present on a catalog. The benchmark conversation ID selects the index during requested access; the records establish no separate identity match selecting delivered memory parts during automatic context selection. Its push-signal unit therefore retains dense, lexical and coarse findings, removes identifier, and remains partial rather than asserting complete coverage. Coarse gates and budgets remain separate from dense/lexical targeting. See RT-RTE-python-invocation, RT-RTE-typescript-hooks, MEM-RTE-fact-selection and MEM-RTE-benchmark-reload.

Trace-fed updates and limits

The ordinary SDK chain qualifies for trace learning because automatic interaction-fed writes produce durable facts/summaries for later model context. Opaque server semantics prevent stronger content/warrant conclusions, not the wired existence of that mediated chain. Raw-log capture by itself, generic-stream persistence without augmentation, benchmark judgment retention and feedback transmission do not supply a qualifying positive or a bounded system-wide negative. Optional host tool traces support a claimed hosted source path, not proof every result reaches extraction. Hook events are triggers, not evidence that the update consumes event-stream content. See RT-RTE-augmentation-admission, RT-RTE-stream-exhaustion, MEM-OBJ-agent-memory and MEM-RTE-feedback-submission.