Agent S memory profile
Type: agentic-system-analyses/types/agentic-system-memory-profile.md
Comparison rationale
The profile follows the accepted records, not an inferred uniform architecture. Historical S1/S2 experience and query caches, current ACI retained context, and offline comparison have different write mechanisms and later consumers. Their local findings remain separate. All axis inventories retain unresolved initialization and provider-state alternatives; partial coverage is not weak implementation evidence.
Content, access and authority
MEM-OBJ-narrative-memory and MEM-OBJ-episodic-memory provide optional advice. Placement in a system prompt does not turn retrieved recipes into standing instructions. MEM-OBJ-embeddings supplies ranking at a numerical consumer, not textual advice or a model-weight update. MEM-RTE-query-cache additionally routes search inputs. MEM-OBJ-judge-results route scoring and enforce a resume skip through task-key presence; their retained rationale is not the field those consumers use.
The raw historical work text has learning authority at summary generation, while the resulting summary has knowledge authority at later planning. The MEM-OBJ-code-result similarly separates the history-to-summary update from advisory delivery. Missing controller-output detail limits warrant, not the existence of this update. MEM-OBJ-traj feed retained captions; MEM-OBJ-fact-captions then provide evidence and ranking influence at comparison. None of those paths establishes improved capacity, correct captions or exercised beneficial self-change.
Transformations and qualifying traces
Summary production is consolidation of retained work, including when the input trace remains. This classification does not certify semantic preservation. The epistemic account's indeterminate content relations concern whether particular propositions are preserved, entailed or conjectured; unseen output cannot establish synthesis. The MEM-RTE-knowledge-save retains afforded findings rather than borrowing the CLI's wired caller. Its caller-supplied original trace mixture remains unknown.
The coding summary qualifies as retained continuation context even within a task. Its input includes recorded model thoughts and action requests, not direct controller-result dictionaries. Offline captioning also qualifies because its automatically produced durable descriptions shape a later comparison invocation. Its source is linked screenshot/action experience and recorded execution actions. Ordinary reflection, raw logging without a transformation, and context trimming do not independently establish trace learning. Generic note strings and cached answers retain unresolved trace origins rather than inheriting the historical summary chain's positive value.
Request and selection boundaries
MEM-RTE-experience-retrieval explicitly fulfills a consumer request. Its cosine ranking and exact-key skipping do not establish an unsolicited embedding-selected push. Query-cache and offline loader returns are also pull. MEM-RTE-notes-delivery and the pending coding return are automatic message-construction pushes: availability selects the full applicable content. They support coarse selection, not identifier targeting merely because a task label appears in the payload.
Canonicalization and limits
Reconciliation supersedes the epistemic caption and comparative-choice object declarations with MEM-OBJ-fact-captions and MEM-OBJ-judge-results. The profile cites those canonical records. Other overlaps are parts or cross-generation scopes, not replacement identities: the S3 code summary remains part of the larger return, and the S2 admission container is not counted as another store beside its content objects.
RT-RTE-kb-bootstrap establishes software import, not original entry authorship, human review or complete archive inventory. Existing vector provenance is likewise unknown. The RT-CMP-lmm-agent, RT-CMP-grounding-model and RT-CMP-embedding-engine component records leave provider-side retention and learning opaque. External model components do not establish a model-weights substrate for use-accumulated memory. These missing facts prevent stronger axis coverage; no local positive witness upgrades those alternatives.
Source-read log: none. This classification used only the supplied accepted members and their governing contracts. No source passages, new records, observed runs or causal evidence were added.