HippoRAG memory profile

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

Comparison rationale

The profile classifies accepted retained parts and consumer paths, not all runtime storage. The default stores separate original text, normalized symbolic strings and dense numerical representations. Graph and vector values describe access structures as well as their file-backed persistence. A stored embedding is not a model-weight update. See MEM-OBJ-passage-embeddings, MEM-OBJ-fact-text and MEM-OBJ-graph-state.

The reconciliation supersedes the duplicate epistemic passage object with MEM-OBJ-passage-text. All passage classifications cite that canonical record. The OpenIE container, extracted relation content, normalized text, vectors and graph remain distinct; no other supersession changes a value. The profile declares no records and adds no source evidence.

Behavioral authority follows each consumer. Text supplies knowledge; stored vectors, facts and graph affect ranking; saved coverage and source mappings route later work; manifest fields provide compatibility criteria and an enforcing veto. Metadata returned for display is not an instruction or an authorization control. Raw response replay supplies prior content to a requesting inference consumer. Neither reuse nor retention itself establishes learning authority. See RT-OBJ-index-manifest, MEM-OBJ-chunk-metadata and MEM-RTE-response-reuse.

Extraction consolidates source material into a reduced search representation. Its semantic reliability is a separate question: the accepted epistemic record leaves entailment indeterminate. That uncertainty does not turn extraction errors into synthesis. Deduplication is supported by explicit triple removal rather than solely by primary-key acquisition checks. Metadata replacement evolves an entry. Physical deletion is forgetting (decay), not invalidation retaining a criticism history, and does not imply cache erasure. See EPI-RTE-openie-generation, EPI-RTE-triple-admission, EPI-RTE-text-normalization and MEM-OBJ-chunk-metadata.

Explicit conditional user approval in WF-RTE-folder-indexing supports claimed manual admission at that checkpoint only. It does not upgrade host adherence to wired or make downstream model/software admission manual. Generic caller identity leaves initial content control unresolved. Exact producer weights and opaque adapter lifecycles also remain unresolved rather than inferred from a default witness.

Requested selection remains pull even where similarity scoring and model judgment automatically select the returned content. Exact request-key replay is likewise not identifier-based push. Alternative cache routes are not completely inspected, so the signal axis is not globally inapplicable. See RT-RTE-retrieve-answer and MEM-RTE-response-reuse.

Trace-learning coverage is deliberately local. Raw response replay and current-call thought accumulation do not establish a qualifying derived trace-fed write. Arbitrary imported document content has no recorded trace-origin inventory; treating it as necessarily static or necessarily a session log would add evidence. No canonical bounded-absence record supports a controlled no finding, so the trace-learning axis remains not-determinable. The trace-source axis cannot inherit a system-wide negative. No observed operation, beneficial learning or causal improvement is claimed.

Source-read log: none. This classification used only the supplied accepted members and reconciliation; no target-source ambiguity required an additional read.