Ultimate memory profile

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

Comparison rationale

Boundary and coverage

This profile classifies accepted records, not a new source inspection. Known units describe the named local mechanisms, not all external implementations or deployed content. Every axis remains partial because the records retain material opaque or less-inspected alternatives. Provider inference alone does not make model weights accumulated memory; provider retention/update internals remain unresolved rather than becoming a positive parametric-memory finding. The local SQLite integration does not certify external transactional correctness (RT-OBJ-agent-doc, RT-CMP-provider-models, MEM-CMP-jev).

Canonical identities

Reconciliation supersedes the epistemic memory object with MEM-OBJ-memory, the goal object with MEM-OBJ-goals, memory consumption with MEM-RTE-memory-read, review with MEM-RTE-review-memory, chapter carry with MEM-RTE-chapter-seed, and the tuning claim with MEM-CLM-recall-tuning. Those identities govern the citations above. The separate epistemic candidate-generation, normalization, checks, retention and relevance records remain distinct functions, not superseded whole routes. No disagreement requires choosing a strongest status.

Force, admission and transformations

Retained memory can supply knowledge to a work model, instructions when its content directs behavior, ranking input to recall, and learning input to a durable update. These are separate consumers, not blanket properties of all text. Existing map membership supplies validation and an enforced admission veto, without turning structural checks into truth acceptance (MEM-BAP-memory-guidance, EPI-RTE-memory-checks). Goal authorization is interpretive instruction; state/counters have separate code effects (EPI-RTE-goal-authority).

Automatic admission is not manual because a user began a task or selected a chapter transition. Human deletion, direct editing and goal authorization are manual paths. Ordinary review affords human choice, while hands-off policy affords computational substitution. Generic extension callers remain unresolved. Seed application is a separate automatic installation route and bypasses ordinary batch validation (MEM-RTE-memory-edit, MEM-RTE-review-memory, MEM-RTE-chapter-seed).

Summary compression supports consolidation without requiring raw-history deletion. Memory overwrite supports evolution. Same-slug seed selection resolves key collisions, not established duplicate content, so it emits no curation value. The separate afforded review-cleanup finding does not establish wired seed deduplication. Bounded display/completed-task retention supports decay. Landing generation remains semantically indeterminate: a model request for lessons neither proves synthesis nor licenses classifying every landing as faithful consolidation. The explicitly inspected summaries still support consolidation independently (MEM-RTE-summary-tally, EPI-RTE-landing-candidates, MEM-OBJ-call-summaries).

Selection and trace-fed writes

Requested memory/history reads are pull; automatic brief, recall, catch-up, operational-note and continuation delivery are push. Recall uses inferred judgment over memory lines; merely returning slugs does not establish a separate identity-match selector. Catch-up combines coarse recency/budget selection with path-overlap targeting. A requested substring history search is not lexical push (MEM-RTE-memory-read, MEM-RTE-recall-message, MEM-RTE-tally-catchup, RT-RTE-codemode-tools).

Landing-derived memory, abstracts and retained continuation summaries support wired trace-learning yes. This classification concerns automatic trace-fed durable guidance, not measured improvement. Session-derived abstracts keep their session origin even inside a landing event; the discrete landing record separately supports event-stream origin. The records do not establish every original tool payload or a trajectory-learning route. No bounded absence record supports a local no, and no opaque alternative is dropped to obtain complete coverage (MEM-RTE-landing-distil, MEM-RTE-summary-tally, RT-RTE-context-return, MEM-RTE-rewind-context).

Call-summary display lines, effect flags, receipts and header fallback retain unresolved lineage as a separate unit; they are not covered by the deterministic catch-up/seed finding. Their prose form at the display consumer is known without asserting a foreground memory route (MEM-OBJ-call-summaries).

Source-read log

No frozen-source reads were made. All classifications use the supplied accepted records and annotations; no new evidence or record is declared.