Stigmergy memory profile

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

Comparison rationale

This profile classifies accepted records, not a live tenant corpus. All positive findings are wired architectural mechanisms. No supplied execution or causal experiment supports observed activation, improvement, or causal benefit. Each partial axis preserves its unresolved included surfaces rather than treating an inspected corpus route as the whole inventory.

Scope and operative parts

MEM-OBJ-artifact, MEM-OBJ-source, MEM-OBJ-knowledge-body, MEM-OBJ-page-metadata, MEM-OBJ-contradiction-data, MEM-OBJ-contradiction-display, MEM-OBJ-fragment, MEM-OBJ-fragment-vector and MEM-OBJ-knowledge-changes retain separate parts and consumers. Binary originals are not classified from readable extraction alone. The remote embedding producer is not the retained vector. The in-process evidence-store alternative is included without a durability claim.

Operational recovery state is included where retained commits/manifests determine later reconciliation, and replay/session state where it directs later consumers. The accepted records do not fully characterize audit/job/report/roster objects or every consumer, so these remain included unresolved units. Static librarian instructions and product release machinery are not accumulated memory merely because stored in the same source repository. Fresh per-question ledgers and draft plans are ordinary invocation state; their delivery does not establish a retained continuation summary.

Agency, transformations and authority

RT-RTE-capture-filing, RT-RTE-maintenance-admission and MEM-RTE-alias-recording establish automatic admission of derived content. A human starting maintenance does not manually approve its generated output. MEM-RTE-workspace-replay separately establishes operator-confirmed import/handover. Master authorization on other APIs does not establish that every caller is human.

EPI-RTE-entity-merge supports consolidation and duplicate removal independently of unresolved proposal semantics in EPI-RTE-filing-transform. Neither a requested synthesis nor an unverified factual change establishes the controlled synthesis value. Ordinary withdrawal under EPI-RTE-evidence-withdrawal invalidates current managed reliance while Git history remains; destructive history erasure has a different horizon. No decay or promotion mechanism is inferred from editorial status labels.

Corpus text supplies knowledge at the later planner/answer consumer. The same text, aliases and vectors supply ranking inputs at selectors. Metadata and canonical conflict payloads supply criteria and vetoes at code gates. These distinct paths do not make all corpus text standing instructions. Retention of mutation reasons in change manifests does not demonstrate an automatic later model reader or learning authority for those reasons.

Requests, automatic selection and limits

MEM-RTE-scoped-reading establishes explicit pull routes. Within RT-RTE-capture-filing and RT-RTE-scoped-answer, code also supplies selected retained context before the model requests those pages. That operation is push even though an upstream capture or question triggered the enclosing workflow. Identity matches, lexical selection, vector similarity and coarse visibility/budget limits retain separate warrants; a requested page identifier is not itself evidence of targeted push.

The accepted memory annotation qualifies the filing route's nominal context cap: compact-size trimming can still fail indented rendering, and one candidate may survive while excess source evidence is removed. This prevents a guaranteed delivery-size or total-relevance conclusion, not classification of the automatic selectors. The answer records disagree about whether every passing citation quote was actually delivered earlier to the model. EPI-RTE-answer-check retains a refetched-head path; RT-RTE-scoped-answer uses surfaced-content wording. No classification here requires snapshot identity or prior delivery of every passing quotation. Scoped retained-content supply and pull tools remain established by both accounts.

Trace inventory and unions

The automatic write and later-consumer links do not resolve original trace input. MEM-OBJ-artifact, MEM-RTE-alias-recording and EPI-RTE-filing-transform leave mixed capture origins and concrete semantic transformations unresolved. Raw queue/audit/evaluation retention is not itself a qualifying trace-fed update. No accepted evidenced-absence declaration covers these alternatives. Therefore neither local trace-learning negative nor positive is established, and trace source is not inapplicable. The derived system union has no supported trace-learning value; this does not claim absent learning or absent improved capacity.

All classifications cite declaring records rather than new evidence. Reconciliation contains no supersessions affecting these values; valid containers and operative parts remain distinct. Its unresolved context-budget and quote-delivery disagreements are retained as limits above, not settled by choosing a stronger status.

Source-read log

No frozen-source reads were made. The supplied records and their retained passages sufficed; no new records, evidence, quotes or report corrections were introduced.