MentisDB memory profile
Type: agentic-system-analyses/types/agentic-system-memory-profile.md
Comparison rationale
The inventory separates content from its envelope, uploaded instructions from their registry, and embedding outputs from inference weights. MEM-OBJ-thought-content, MEM-OBJ-skill-body, MEM-OBJ-embedding-vector and MEM-OBJ-retrieval-state carry those distinctions. The compiled core skill and provider models are not accumulated memory merely because they guide or compute a memory operation. No observed-run or causal basis is asserted.
Reconciliation supersedes the epistemic duplicate thought-content, skill-document, summary-candidate, relation-withdrawal, summary-selection, bundle-selection and recall-evaluation records with their memory declarations. This profile cites the canonical memory IDs. These supersessions preserve the original findings; they do not upgrade consumer affordance to wired summarization. The Markdown import subroute and skill-withdrawal subroute remain distinct accepted epistemic records; only the latter is needed here.
The admitted content, actual update consumer and later consumer determine authority. Retained text influences ranking even under fixed code. Default invalidation has an enforcing selector; revoked skill content remains readable, so lifecycle retirement does not enforce host nonexecution. Structural reference/hash checks concern conformance, not factual truth. MEM-RTE-correction-update, EPI-RTE-skill-withdrawal, RT-RTE-memory-roundtrip and WF-RTE-session-continuation bound these different effects.
Raw conversation storage is not itself trace learning or consolidation. The positive trace-learning finding belongs only to the automatic numerical derivative: recorded messages can become durable vectors and neighbor state, and later search consumes that derivative. Its update-consumer learning force and downstream ranking force are separate. This classification requires neither a new proposition nor measured improvement. It does not classify the optional extraction response as retained learning, infer model training, or call a Summary-labelled raw message a summary. MEM-OBJ-thought-content, MEM-OBJ-embedding-vector, MEM-RTE-derived-update and MEM-RTE-retrieval-evaluation establish the input/derivative/later-consumer chain and its limits.
Caller-authored semantic summaries and skill revisions remain separate alternatives. Their documented semantic consolidation is afforded; the source-window selector neither generates nor persists it. Generic caller identity does not resolve human approval versus automatic admission, so no manual agency or automatic semantic trace-learning finding follows. Initial skill provenance and mixed derivative inputs remain unresolved even where compilation is known. MEM-RTE-summary-selection, RT-RTE-skill-admission and MEM-OBJ-embedding-vector retain those missing facts.
Deletion is explicit forgetting, not timed decay or guaranteed secure erasure. Correction keeps history while changing default reliance; skill evolution preserves earlier versions. Optional lexical supersession provides a wired near-duplicate removal from current retrieval, without checking equivalent meaning. These are distinct curation operations, not names inferred from index rebuilding. MEM-RTE-chain-withdrawal, MEM-RTE-correction-update and RT-RTE-skill-admission support the mappings.
Known requests are pull. Automatic ancestor traversal and Python formatting fulfill a request, rather than adding push operations. The webhook selects a new append event, not accumulated historical memory. Unknown later diagnostic/log/callback consumption prevents a complete pull-only boundary and therefore prevents an axis-wide inapplicable push signal. Missing facts stay unresolved rather than becoming bounded negatives. RT-RTE-memory-roundtrip, MEM-RTE-webhook-delivery, MEM-OBJ-retrieval-state and MEM-RTE-retrieval-evaluation supply those limits.
Source-read log: none. Classification used only the supplied accepted members and reconciliation; no source ambiguity required an additional frozen-source read.