Eigenius memory profile

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

Comparison rationale

The classifications concern retained material and its actual later consumers, not every storage or permission mechanism in the runtime. All positive implementation findings are wired; manual notebook editing/publication is an afforded operator route. No accepted record supplies observed operation or causal benefit. Partial axes retain the included uninspected alternatives rather than allowing one implemented witness to stand for a complete system inventory.

Objects, representations and provenance

The layered resource graph, result-key metadata and vector access segments have different storage roles. Graph structure and numerical vector storage can coexist in the HNSW access path. Provider weights are excluded: the accepted records establish inference dependencies, not parameters accumulated or changed through this platform's use. Dense segments are parametric representations, not learned provider weights. Symbolic JSON decision fields remain distinct from linguistic rationale and gloss fields. A readable rendering does not replace the encoded consumed artifact. See MEM-OBJ-vector-segment, MEM-OBJ-selection-record and RT-OBJ-formalization-artifact.

Trace-extracted lineage is assigned to automatic live-exchange recording; externally supplied recordings have imported lineage at entry. The same object kind can therefore have different derivation paths. Generated resource artifacts and index segments have other-compiled lineage without a claim of new knowledge. Initial lexicon, caller and embedding provenance remain unresolved. Published notebook serialization does not establish that every cell was human-authored. See MEM-RTE-draw-retention, MEM-RTE-index-refresh and MEM-RTE-notebook-reopen.

Authority and maintenance

Retained decisions route or rank parser work, retained completion strings instruct a model, and retained graph/vector material supplies knowledge or retrieval ranking inputs. Formal witness evidence instead has validation and enforcing force at a specific checker. It does not grant generic truth authority to all retained assertions. The recorded experience has learning authority at production of the durable replay artifact, separately from that artifact's routing/ranking effect at replay. See MEM-BAP-recorded-decision, RT-BAP-completion-prompt and EPI-RTE-witness-consumption.

Layer consolidation is reduction of the effective representation of existing definitions, not semantic synthesis. Tombstones and witness-credit withdrawal invalidate current reliance; eligible unreachable-layer deletion is decay in the storage-forgetting sense. Vector segment replacement under changed model declarations is access-index rebuilding alone, not a qualifying evolve operation. Its bounded curation unit is inapplicable; no separate memory-entry revision is established by the accepted record. No dedup value is inferred merely from content addressing or primary-key identity, and no synthesize value follows from unassessed model/query output. These classifications do not certify erasure, crash consistency, scientific warrant or complete maintenance coverage. See MEM-RTE-state-maintenance, MEM-RTE-index-refresh and EPI-RTE-witness-consumption.

Read-back and trace boundary

A query, notebook reopen or artifact fetch returns requested retained material: pull. A requested formalization can also contain a subsequent automatic keyed replay selection: push at the decision consumer, not a second push of the caller's requested result. Post-commit installation of declaration-derived dispatch state is coarse push. A witness-dependent check selects actual retained evidence by target/proposition identity. Pull-query lexical/vector ranking is not imported into the push-signal axis. See MEM-RTE-query-resources, MEM-RTE-draw-retention, MEM-BAP-institution-refresh and EPI-RTE-witness-consumption.

The automatic decision-exchange recorder has a durable later replay consumer and therefore supports a local trace-learning yes without parameter training or demonstrated improvement. It does not establish any controlled original trace-source category: document prose and model decision exchanges are not automatically conversational session logs, tool traces, environmental event streams or trajectories. That missing origin fact leaves trace_source not-determinable. Other unresolved routes are not made inapplicable by the known replay route. See MEM-RTE-draw-retention, MEM-OBJ-selection-record and RT-RTE-program-execution.

Replay's horizon is explicitly bounded. Branch recreation plus recording only live exchanges can remove old decisions from the next default branch after an all-replay invocation. This limits persistence coverage without withdrawing the implemented subsequent-consumer path. Sense/proposal rationale use, original generic-caller control, vector input provenance and recovery remain independent gaps. See MEM-RTE-draw-retention, MEM-OBJ-rank-record and MEM-OBJ-proposal-record.

Reconciliation and source-read log

The supplied reconciliation contains no supersessions and no unresolved cross-report conflicts. All citations therefore retain their declaring records. The broader query route and narrower FIBER retention route remain valid overlapping scopes; neither replaces Load or program execution. The profile does not merge domain proof inputs, generated artifacts and general corpus containers into one provenance chain. See MEM-RTE-query-resources, EPI-RTE-fiber-retention and EPI-RTE-lean-check.

Source reads for this classification: none. Only the supplied invocation, types, contracts and accepted members were read. No new evidence or records were introduced, and no report verification was reopened.