A-MEM memory profile

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

Comparison rationale

The comparison concerns retained note parts and their access structures, not all runtime state. Instance-local accumulation is sufficient for later read-back. It does not imply restoration across construction: the map starts empty and initialization attempts to reset the collection. The collection's dense vectors, readable enhanced text and symbolic metadata remain separate forms. A link list is a symbolic access structure in Python/Chroma, not evidence of a graph substrate (MEM-OBJ-chroma-snapshot, MEM-OBJ-note-links, RT-RTE-note-admission).

Reconciliation supersedes the epistemic note-content and semantic-metadata identities with MEM-OBJ-memory-content and MEM-OBJ-semantic-metadata. Those canonical objects are cited here. The enhanced-document part remains distinct from its valid Chroma container, so its natural-language classification cites EPI-OBJ-enhanced-document. No route supersession changes a value.

Automatic evolution establishes entry revision because code replaces existing tags/context. It does not establish semantic synthesis or faithful consolidation. The models are asked for summaries and reinterpretations, but actual content relations remain indeterminate. Index consolidation merely republishes representations. Caller deletion removes retained entries from the retriever and live map, supporting decay as forgetting without requiring gradual downweighting or automatic selection. It does not retain withdrawal history, so it does not establish invalidate; surviving links may dangle and previously delivered copies remain outside this deletion mechanism. Thus evolve and decay are supported without importing operations suggested only by method names (EPI-RTE-evolution-disposition, EPI-RTE-evolution-judgment, MEM-RTE-index-consolidation, RT-RTE-direct-mutation, MEM-OBJ-note-links).

Retained descriptions have knowledge authority at later evolution and reader consumers, and ranking authority when enhanced text determines similarity. Link IDs route expansion. Call-local model decisions authorize mutations, but are not themselves retained standing controls. A later add's automatic embedding selector pushes neighbors to the model; explicit retrieval pulls material to the caller. Requested return is not another push. Neither delivery nor mutation establishes actual reliance, semantic correctness or demonstrated benefit (RT-BAP-neighbor-context, RT-BAP-caller-results, MEM-OBJ-note-links, RT-RTE-memory-retrieval).

Unknown initial input provenance and generic caller identity are consequential. They prevent authored/manual classifications and prevent assigning a trace source. The automatic transformation chain is not erased, but cannot establish trace learning without trace-origin evidence. Conversely, the bounded absence of usage/history feedback cannot establish a global negative while caller-provided histories and provider processing remain unresolved (MEM-OBJ-memory-content, MEM-OBJ-usage-fields, MEM-ABS-usage-feedback, RT-RTE-direct-mutation).

Provider-side memory is retained as an unresolved alternative, not asserted to exist. Static inference weights are excluded from the accumulated-memory inventory; their unchanged local parameters do not answer trace learning through artifacts. All positive bases are wired implementation, not observed operation or causal benefit. No source reads were made for this profile; classification used only the supplied accepted members and reconciliation.