Engraph memory profile

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

Comparison rationale

This profile classifies accepted records, not new source facts. It includes retained source content, derived access/selector state, write feedback and unresolved alternatives. Model weights and shipped instructions are excluded because these records do not show their accumulation or change through Engraph use. Current-call bundles and health reports are not retained objects; their retained inputs and later consumption routes remain included. The inventory and limits are supported by MEM-OBJ-vault-notes, MEM-OBJ-chunk-record, MEM-OBJ-vault-profile, MEM-OBJ-health-report, and MEM-OBJ-cli-event.

SQLite is the substrate for the graph, FTS and vector structures. Their encodings and access algorithms do not imply separate graph/vector storage services. Dense chunk embeddings and centroids are parametric representations, distinct from the excluded loaded model parameters. Their provenance is only partially resolved, even though their storage and form are known (MEM-OBJ-chunk-record, MEM-OBJ-folder-centroids).

Reconciliation identifies the epistemic identity-fact, identity-extraction and placement-correction declarations with the canonical memory objects/routes cited here; the health-output duplicate likewise resolves to the memory object. Those amendments change citation destinations, not values or evidence strength. They preserve the identity extraction counterexample, source loss and activity/importance promotion. Thus extraction includes both consolidation of existing text and synthesis of non-entailed category associations. Faithfulness loss in copied focus/blocking bullets alone would not establish synthesis. The canonical route remains MEM-RTE-identity-context, and the numerical feedback route remains MEM-RTE-placement-feedback. No other supersession changes these classifications.

Move feedback establishes automatic trace-fed durable behavior shaping: an environmental move plus retained vectors changes centroids consumed by later writes. Old vectors have learning authority at the update; updated centroids rank and route future placement. The raw correction log is not the inspected placement input. Event-stream origin is supported for the move, not for every original note that contributed vectors. The inconsistent centroid writer keys/counts prevent a uniform full-vault mean or complete reset inference, not existence of the update pathway (MEM-RTE-placement-feedback, MEM-OBJ-folder-centroids, MEM-OBJ-placement-corrections).

Requested note/identity/search/context returns are pull. Their automatic internal supply of cache, profile, naming and placement state is a separate operation, not a second push for every returned search hit. Cache query-hash equality actually selects retained JSON; a requested note identifier alone does not establish identifier push. Automatic extraction supplies selected retained snippets to an update consumer rather than unsolicited identity text to an agent (RT-OBJ-llm-cache, RT-RTE-query-search, RT-RTE-index-refresh, MEM-RTE-vault-readback, MEM-RTE-identity-context).

Agency remains partial because caller identity does not establish human per-content decisions. Starting a workflow does not turn tag/link/placement or identity extraction into manual admission. No manual finding follows from the CLI label or the opportunity to review a migration preview (RT-RTE-note-admission, RT-RTE-setup-apply, RT-RTE-para-migration).

Curation distinguishes retained changes from transient formatting or indexing. Archive can invalidate ordinary indexed reliance while preserving a body; hard deletion forgets selected material, but soft deletion and auxiliary cleanup do not support complete erasure. Feedback subtraction downweights one numerical representation, not time-based forgetting. CLI-event age pruning is an afforded method with untraced activation, not a wired maintenance loop (MEM-RTE-note-withdrawal, MEM-RTE-placement-feedback, MEM-OBJ-cli-event).

Unknown telemetry consumers, arbitrary source provenance/control, uninspected graph-output details and unresolved lifecycle derivatives prevent stronger coverage assessments. No bounded absence records establish local trace-learning no or a pull-only boundary. Unknown source origins do not justify invented session/tool/trajectory values. Every positive remains at its own wired or afforded basis; no observed operation, causal benefit, client obedience or improved placement is asserted.

Source-read log: none. Classification used only the supplied members, types, collection and records contract; no source ambiguity required a frozen-checkout read.