sqlite-graph-memory memory profile

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

Comparison rationale

This profile classifies accepted records, not target execution. Wired findings describe implemented paths; afforded inspection describes an available consumer role, not a deployed dashboard or an observed maintainer decision. No finding claims observed activation, improved capability or causal benefit.

Retained boundary and representation

The package name does not make its note index a SQLite graph. Source Markdown retains links; NumPy vectors and pickled snippets are separate files, while SQLite holds transcript and comparison records. Graph and vector findings name access structures over retained material, not SQL schema claims (MEM-OBJ-notes, MEM-OBJ-brain-e5, MEM-OBJ-brain-meta, MEM-OBJ-turns, MEM-OBJ-ab-recall).

The numerical index is parametric representation, not updated encoder weights. Natural-language snippets and symbolic path/date fields have separate consumer paths even though one pickle holds both. In-memory candidates, PPR cache and the benchmark raw-note map are working copies or invocation state, not another accumulated store. The benchmark can nevertheless consume retained source-note text beyond the snippet prefix (MEM-OBJ-brain-e5, MEM-OBJ-brain-meta, MEM-OBJ-benchmark-raw, RT-RTE-recall-options).

Admission, lineage and transformations

External note origin stays unresolved beside known import and automatic index derivation. A caller starting indexing does not supply a per-note human approval decision. Similarly, evaluator reference inputs do not reveal their original admission control. No manual finding is licensed merely by operator installation or a CLI trigger (MEM-OBJ-notes, RT-RTE-index-admission, EPI-OBJ-evaluation-records).

Consolidation identifies reduction of retained text: transcript extraction and snippet clipping. Duplicate removal in captured lists supplies a separate deduplication finding. Changed-file refresh revises existing index entries; it does not preserve withdrawal history, certify a corrected claim or synthesize new knowledge. Temporal closing and reason retention remain external-writer prescriptions, not an implemented package invalidation operation (MEM-OBJ-turns, RT-RTE-turnstate-capture, RT-RTE-index-admission, RT-BAP-temporal-instructions, RT-CLM-temporal-scope).

Consumer force and read-back

Snippet delivery supplies knowledge context, while retained vectors, text and links supply ranking inputs. Manifest and offset metadata route subsequent maintenance work. They do not become instructions merely because they affect behavior. Stored evaluation references and hashes supply validation criteria; the configured result gate can enforce rejection locally. These forces do not validate note truth or establish system-wide merge enforcement (RT-BAP-context-delivery, MEM-OBJ-brain-manifest, MEM-OBJ-turnstate-offset, EPI-RTE-bench-score, EPI-RTE-bench-gate).

Requested recall remains pull throughout candidate selection and delivery. Graph expansion is not an unsolicited context injection. Automatic control metadata supply is different: index-build code matches prior manifest entries by path/settings, and Stop-hook code selects a session-specific offset. These identifier push findings concern control metadata, not semantic targeting of delivered note snippets. Pull-route query embeddings, ledger recency and lexical benchmark extraction do not supply additional push-signal values (RT-RTE-mcp-recall, RT-RTE-index-admission, MEM-OBJ-brain-manifest, RT-RTE-turnstate-capture, MEM-OBJ-turnstate-offset, EPI-RTE-extract-answer).

Trace-learning limits and canonical identity

Automatic transcript extraction reaches a ledger viewer, not an established later behavior-shaping artifact consumer. Offsets select a future input window but are routine checkpoints; telemetry and evaluator history do not establish automatic policy adaptation. Original note trace origin remains unresolved. The accepted records contain no canonical evidenced-absence declaration warranting a local trace-learning "no". Therefore neither "yes" nor "no" is asserted, and trace source remains not-determinable rather than inapplicable. This preserves the distinction between trace-extracted lineage and a qualifying trace-learning route (MEM-OBJ-turns, MEM-OBJ-turnstate-offset, MEM-OBJ-notes, MEM-OBJ-ab-recall, EPI-OBJ-evaluation-records, RT-RTE-evaluation-paths).

Reconciliation supersedes the epistemic turns declaration with MEM-OBJ-turns, so ledger findings cite that canonical object. It also supersedes the epistemic diagnostic route with MEM-RTE-vault-health; those diagnostics do not retain a report history or automatically curate memory and add no profile operation. Neither supersession changes a controlled value. The broader epistemic note/snippet and recall-state containers remain valid overlaps, not aliases of one memory object.

No frozen-source read was needed or performed for this classification. All warrants remain in the accepted record citations; no new records, quotations, execution evidence or amendments are introduced.