claude-obsidian memory profile

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

Comparison rationale

This classification uses the supplied accepted records and their annotations only. No source files were read, no new evidence or records were added, and no target program was executed. Evidence is implementation or declared procedure, not observed use or causal benefit. The package-owned memory inventory includes computational scheduling and recovery state as well as prose; it does not classify every host/provider storage surface.

Parts and consumers

MEM-OBJ-chunks separates natural-language ranking input from symbolic postings and metadata. MEM-OBJ-embed-cache and MEM-OBJ-tiling-cache separate dense parametric representations from keys and thresholds. File encoding is their substrate, not their representational form. Initial embedding provenance remains unresolved even though inference-derived cache writes are wired. Raw files in MEM-OBJ-source-ledger have a known substrate but can have unresolved operative form; locators do not classify their opaque payloads.

Knowledge at later models is distinct from ranking at retrieval code, routing at filing/queue consumers and validation/enforcement at admission or recovery. EPI-RTE-ledger-check uses retained support metadata as criteria; EPI-RTE-ledger-disposition gives violations an enforced veto. Neither establishes semantic truth. RT-OBJ-transaction-bundle, with its memory annotation, supplies retained recovery state, not a second knowledge ledger. No learning authority is inferred merely from reuse of vectors or stored results.

Agency, transformations and delivery

Manual findings describe explicit operator decisions prescribed by RT-RTE-save-apply, EPI-RTE-canonical-integration and the configuration/view/fold procedures. They do not turn the approval hash into an authenticated human judgment. RT-RTE-direct-capture leaves caller control unresolved; code-only derived-cache admission is independently automatic. Initial adoption and arbitrary queue submissions retain their separate uncertainty.

Synthetic prefix copying supports consolidation locally through EPI-RTE-synthetic-prefix. Model prefixes do not inherit that classification, and model synthesis labels do not establish a novel claim. Deleting orphan vectors/surplus chunks is forgetting; withdrawing the BM25 index before mutation is invalidation of an access representation. MEM-RTE-chunk-maintenance and MEM-RTE-tiling-diagnosis do not establish complete semantic withdrawal of source/claim prose.

RT-RTE-vault-query is a requested read chain, including ranking and return. It is not converted into push merely because code selects pages. RT-RTE-session-context separately supplies fixed hot context on lifecycle triggers and emits retained recovery status on Stop. These are coarse selection operations with availability, consent and budget filters; neither query relevance nor a matching content identifier is established. Actual external host delivery and obedience remain unobserved.

Supersessions and unresolved coverage

The supplied reconciliation supersedes the epistemic claim-ledger declaration with MEM-OBJ-claim-ledger; all profile claim-ledger findings use that canonical object. It also supersedes the epistemic extractive-fold declaration with MEM-RTE-log-fold. The fold retains the disagreement between claimed and afforded implementation status: the profile uses claimed for the shared requested extractive transformation and afforded only for its recorded discovery/read path. The curation unit remains partial rather than assigning one unqualified implementation conclusion. No source reread or adjudication reopens report verification.

The source-ledger and derived-context containers are not whole-record supersessions of their memory parts. Separate rerank/tiling use identities remain separate without asserting two deployed models. The profile therefore keeps local consumer warrants rather than upgrading all branches from one executable witness.

No accepted bounded-absence record establishes trace-learning no. Save and folds have recorded trace-derived content, but their reviewed retention does not establish an automatic trace-fed write. Automatic prefix/vector caches consume pages whose initial and mixed lineage remains unresolved. Queue/journal retention alone is not a qualifying behavior-shaping update. These limits in RT-RTE-save-apply, MEM-RTE-log-fold, MEM-OBJ-wiki-pages and MEM-RTE-queue-resumption prevent both a qualifying positive and a bounded negative. Trace source consequently remains not-determinable, not inapplicable. External host compacted summaries are deliberately outside this package-owned memory scope, not opaque alternatives silently dropped to make these axes known.