OpenWiki memory profile
Type: agentic-system-analyses/types/agentic-system-memory-profile.md
Comparison rationale
The comparison covers accumulated memory and later consumers, not all runtime storage. MEM-OBJ-wiki-page is the canonical page object: reconciliation supersedes EPI-OBJ-wiki-markdown with it. No other identity amendment changes a value. The mixed epistemic Claim container remains valid alongside MEM-OBJ-claim, MEM-OBJ-evidence and MEM-OBJ-page-claims; its existence does not justify combining their forms or authorities.
Positive findings retain source-level wiring or the separately afforded middleware admission contract. Requested semantic curation behavior is not an emitted operation. None is observed or causally supported. RT-RTE-native-generation supplies automatic local admission without per-page approval. RT-RTE-host-generation leaves human control unresolved. RT-RTE-maintenance-admission separates local retention from Git merge: an automatically created PR does not establish manual merging. Optional Git retention is not a substitute for the operative filesystem classification.
The same retained page can supply knowledge to a host and ranking features to the lexical selector. RT-RTE-wiki-retrieval, with its memory annotation, establishes both, plus search-only deprecated suppression. Requested search remains pull; BM25 does not become a push signal. Conversely, retained RT-OBJ-page-guidance supplies instructions at a new worker, and MEM-RTE-claims-reconciliation supplies automatically selected issue Claims. Those assignment operations have identity and issue-status selection, independently of explicitly requested full Claim inspection.
Curation follows transformations. EPI-RTE-source-projection reduces precise resources to file-level provenance and deduplicates it. RT-RTE-page-translation revises retained pages but does not establish faithful semantics. MEM-RTE-chat-continuation deletes older database rows, not necessarily older turns in surviving state. Guidance on MEM-RTE-trace-documentation requests patterns, supported implications and incremental refinement, but does not establish creation of a claim absent from its inputs or an implemented revision of existing documentation. These two units retain no positive curation findings; no synthesize value is supported, and evolve remains supported only on the independent Claim-revision and translation routes. The separately wired trace-fed document-write/later-context chain still qualifies for trace_learning, without implying improved capacity, theory criticism or self-improvement.
MEM-OBJ-conversation-history and MEM-RTE-history-offload establish a durable file channel independently of opaque SQLite graph payloads. Fresh worker graphs and page/Claim snapshot restoration do not erase that channel. The unexecuted test's readable text does not resolve deployed payload form, automatic trace-summary production, reload selection, reasons or cross-worker consumption. Those included alternatives therefore remain unresolved on the relevant axes; one known document chain cannot classify them by analogy.
Benchmark state is a later-consumer surface, not a production revision oracle. EPI-OBJ-ledger-verdicts, EPI-RTE-ledger-judgment and EPI-RTE-ledger-reporting support interpreted judgments, fixed reporting and continuing evaluation scope. MEM-RTE-ledger-evaluation supplies continuing artifact/carry wiring but leaves optional persistence and detailed evaluator alternatives incomplete. MEM-RTE-evaluation-transfer supports cache-selected copying before a later host invocation, not observed host reliance or automatic prompt injection of every copied page.
No axis uses an unsupported complete-system absence. Static managed host guidance and shipped skills are excluded because they are not accumulated or changed through use, as distinguished by RT-BAP-managed-guidance and the memory account. Native provider parameters are not an inspected use-derived memory update; RT-CMP-chat-model leaves provider-side mutability uninspected. This exclusion does not claim that providers or external hosts have no memory.
Source-read log: no direct source reads. Classifications use only the supplied accepted records and their annotations; no additional evidence or records were added. Named uncertainty follows missing recorded facts, not missing execution alone.