ContextPilot memory profile

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

Comparison rationale

This profile classifies the handed accepted records, not target execution. All positive findings are wired implementation witnesses. No supplied record establishes observed use or causal benefit. Local uncertainties concern missing provenance, incomplete downstream tracing and unconstrained content generation; they do not cancel the inspected write and consumption paths.

Objects, consumers and boundaries

The four keyed stores are different operative objects, not alternate names for raw history. Their summaries reach later model calls automatically, while full primary bodies are requested by key. Rich-memory reads also select bounded neighbor displays. Dense vectors are retained access metadata and do not make the dictionaries a vector database. Their producing provider's unpinned weights do not erase the known derived-vector path, but they leave initial provenance unresolved. See MEM-OBJ-memory-indexes, MEM-RTE-inference-stores and MEM-RTE-training-stores.

Retained replacement text supplies knowledge/context at the policy consumer. Its symbolic mask separately enforces visibility at the renderer. Standing shipped guidance is excluded from accumulated memory; the records do not grant saved facts special instruction priority. Fixed neighbor scoring gives retained text/features ranking authority without requiring a learned ranking policy. See RT-OBJ-context-masks, RT-BAP-context-policy and MEM-OBJ-memory-indexes.

Training snapshots supply scored/masked experience to the actor-update consumer, which establishes learning authority there. Ancestry and masks also route credit and exclude rows/tokens. Actor state is the retained output of that learning path. Restoring it for continuing training does not independently establish learning authority at the loader: the records do not show that the restored parameters are consumed as examples, experience or feedback for another update. Its downstream authority therefore remains not determinable rather than a second learning finding. See MEM-OBJ-trajectory-snapshots, RT-CMP-actor-model, RT-RTE-actor-update and RT-BAP-reward-actor.

Terminal scoring compares answers against supplied references and fixed format rules. Those criteria are not established as accumulated or changed through use; consequential scoring does not make them memory. They remain outside this profile's memory boundary and supply no validation-authority finding. This does not erase the learning authority of the scored experience at the update consumer. See EPI-OBJ-answer-evaluation, EPI-RTE-terminal-evaluation and RT-RTE-actor-update.

Transformations and selection

Consolidation describes the implemented reduced-history view, not a guarantee of semantic fidelity. Retaining a substring after dropping a qualifier does not establish synthesis. Optional merge combines existing bodies and loses rich fields; its configurable handler is wired, but ordinary default exposure and observed use are not established. Unrestricted new store arguments have no complete semantic classification. See RT-RTE-context-edit, MEM-RTE-training-context, EPI-OBJ-memory-content and EPI-RTE-memory-merge.

Push signals name actual selection operations. All-summary availability is coarse; matching a retained replacement to its message ID is identifier-based. Requested primary reads are pull even though their key is an identifier. Neighbor expansion is an additional automatic push with fixed-feature and optional embedding signals. Recovery reads use availability/recency, not an established relevance judgment. Training ancestry matches select descendant credit; exact additional scheduler ranking remains uninspected. See MEM-RTE-inference-stores, MEM-RTE-training-stores, RT-RTE-context-edit, RT-RTE-training-rollout and RT-RTE-actor-update.

Learning coverage and unresolved alternatives

Automatic trace-fed replacement writes qualify when their result reaches a later invocation, even within the same QA trajectory. The actor-update route also qualifies, independently of whether improved capacity was measured. Synthetic recovery has two branches: a trainable forced-action suffix or conditioning only. Both alter later context; the latter does not supply a trainable synthetic action. Session content, tool feedback and multi-step rollouts retain their different origins. No event-stream value is inferred merely from code-authored cleanup. See MEM-RTE-training-context, MEM-OBJ-trajectory-snapshots and RT-RTE-actor-update.

No evidenced-absence records were supplied. This profile therefore asserts no local trace-learning negative and no absence of additional controlled values. Known learning routes support the derived system union's "yes"; incomplete coverage prevents a complete-value conclusion. A persisted inference trace without a later semantic consumer is unresolved, not a negative learning witness. Initial content control is similarly unresolved rather than manual. See RT-OBJ-full-history, RT-RTE-document-qa, MEM-RTE-training-stores and RT-RTE-actor-update.

Reconciliation supplies no supersessions. Its container/part dispositions leave the combined inference-content account and the four specific store objects valid; training token snapshots remain distinct from inference history snapshots. No value depends on replacing a declared record. The profile cites only the permitted runtime, memory and epistemic declaring members.

Source-read log

No target-source files were read for this classification. Only the supplied invocation, job/type/contracts and handed report members were read; no named record ambiguity required a permitted source reread.