llm-wiki-cli memory profile
Type: agentic-system-analyses/types/agentic-system-memory-profile.md
Comparison rationale
This profile classifies accepted records, not execution results. All wired findings denote inspected capacity. Afforded agent transformations remain separate from wired persistence; no finding establishes compliance, semantic faithfulness, improved capacity or causal benefit. No additional source files were read.
Identity and canonical records
Reconciliation supersedes the epistemic Practice-items object with MEM-OBJ-practice-items, and the epistemic temporal-feedback route with MEM-RTE-temporal-cycle. This profile uses those canonical destinations. The narrower EPI-RTE-practice-grade remains distinct from MEM-RTE-practice-assess, particularly for override history. No supersession changes a comparison value or promotes an evidence layer.
Content, control and consumers
SQLite stores prose and localized control fields together, but their forms and effects differ. MEM-OBJ-derived-indexes supports symbolic lexical structures, not embeddings. MEM-OBJ-book-blocks separates readable normalized text from opaque original bytes; those bytes remain unresolved rather than inheriting the display text's classification. Replication is another route for copies, not enough evidence to assign a service-object or graph substrate to every opaque component.
Knowledge at a reference consumer does not become instruction merely because pi appends it to a system prompt: RT-BAP-wiki-context. Soul and purpose/schema have a different standing-guidance role. Stored weights and text can have ranking force without retained ranking policy. Stored lease/revision/source-state criteria can validate and enforce admission without establishing semantic truth. The WF-RTE-ingest-preparation packet consumes retained purpose/schema and bounded summaries; its current job-state bookkeeping is not automatically trace learning.
Admission and curation
Generic callers may be agents or humans. A software transaction does not settle who decides the retained content. Accordingly, known automatic capture/scheduler mechanisms coexist with unresolved generic submission control. Sensitive Soul and destructive archive approval are classified only at the claimed policy layer, not as authenticated human approval on all implemented paths. See RT-RTE-soul-publish and MEM-RTE-archive-admit.
Page replacement supports evolve; temporal eviction supports decay; Discussion dependency flags support invalidate. Summary routes afford consolidate. Neither the name Book synthesis nor a missing proof of entailment establishes a novel claim, so synthesize is not asserted. Content-hash reuse and index rebuild likewise do not establish semantic dedup. The semantic uncertainty is retained beside MEM-RTE-book-synthesis, not resolved by changing a route's name.
Delivery and trace-fed updates
Book windows and Practice queues fulfill explicit requests even though software chooses a range or orders due items. They remain pull. Tag pages are independently supplied at lifecycle events, while tracked Plan/Todo cues use a real context match. Budget omission is not inferred lexical or semantic selection. Non-pi and no-context fallback alternatives remain unresolved, as RT-RTE-hook-context and MEM-RTE-task-continuity require.
The strongest trace-learning witness is review rating/time → FSRS card/debt → later queue: MEM-RTE-practice-review. Its original input is a discrete review event. Discussion summaries and teaching checkpoints afford conversational-trace continuation at a different evidence layer. Automatic raw reply capture alone supplies no derived learning artifact. Soul, learner facts and other experience-mediated write-back remain unresolved where original input, control or consumer facts are missing. No absence record warrants a local trace-learning no; the supported yes does not resolve that remaining inventory. See MEM-RTE-discussion-capture, MEM-BAP-discussion-summary and MEM-OBJ-tutor-turns.