Nuggets memory profile

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

Comparison rationale

The profile separates readable content, localized control fields and reconstructed dense numerical state. HRR is distributed-parametric form on an in-memory substrate, not a vector-service substrate or provider-weight update. Its construction and later ranking use are supported by RT-CMP-hrr-vectors, EPI-RTE-recall-decode and MEM-RTE-recall-selection. Fact enumeration and hydration also consume the stored strings directly.

Automatic admission is independent of human input authorship. Model remember proposals, extraction hooks, operator-launched migration and threshold promotion all have software/model admission without per-content approval. Generic caller and initial-file controls remain unresolved, rather than becoming manual by default. See MEM-RTE-fact-mutation, MEM-RTE-hook-capture, MEM-RTE-legacy-migration and RT-RTE-memory-promotion.

Authority is consumer-relative. Fact/path text is knowledge at the model, retained preferences supply instruction, and experience supplies learning at the automatic extraction consumer. Stored keys/values and numerical state can rank recall even though the scoring algorithm is fixed. Retained counters validate and enforce promotion eligibility, not factual truth. Scheduling controls route/gate a later invocation; the retained prompt supplies its instruction. These findings do not imply epistemic acceptance or improved capacity. See RT-BAP-facts-prompt, MEM-OBJ-hit-metadata, RT-OBJ-cron-jobs, RT-BAP-scheduled-prompt and EPI-ABS-truth-acceptance.

Requested recall is pull despite automatic internal decoding. Branch-detail replay and per-turn formatting are stages of an unsolicited push chain: tool-result identity selects snapshots, while category/availability/order budgets select injected content. Cron clock matching is coarse push selection; recipient and job identifiers used for routing/metadata are not, by themselves, identity matches selecting delivered memory. The Markdown writer does not establish its external loader's direction or signal. See MEM-RTE-context-rebuild, MEM-RTE-recall-selection, RT-RTE-proactive-turn and MEM-OBJ-promoted-memory.

Trace learning is a route classification, not demonstrated improvement. Preference/path extraction and retained task excerpts have automatic writes and later context/recall consumers. Session-marked recall also writes durable usage metadata that determines later promotion placement. Ordinary extension recall omits the enabling session ID; this route's existence does not establish activation. Tool paths copied only locally before compaction have conditional durability, so that alternative remains partial. Input-message content is session-log origin, tool-call paths are tool-trace origin, and recall access-counter inputs are discrete access events. No linked state/action/outcome sequence establishes trajectories. See MEM-RTE-hook-capture, MEM-RTE-compaction-extract, MEM-OBJ-task-summary, MEM-OBJ-hit-metadata and EPI-RTE-recall-score.

Curation findings follow implemented transformations. Selected text reduction is consolidation, including where extraction loses qualification; that defect does not establish synthesis. Optional oldest-entry eviction is decay, overwrite is evolution, and threshold placement in exported context is promotion. Primary-key merge alone does not establish deduplication. Forget/delete lacks retained withdrawal history and therefore does not warrant invalidate. Independent unretracted snapshots and Markdown copies also prevent a global withdrawal claim. See EPI-RTE-preference-capture, MEM-RTE-fact-mutation, MEM-RTE-context-rebuild and RT-RTE-memory-promotion.

Reconciliation supersedes the epistemic task-object declaration with MEM-OBJ-task-summary, and the epistemic migration-route declaration with MEM-RTE-legacy-migration. This profile uses those canonical records; supersession changes citations, not the retained extraction, lineage or consumer limits. The residual dispute about memory's attribution of older runtime wording changes no classification: the current RT-RTE-memory-context and MEM-RTE-compaction-extract both slice the final twenty tool-use blocks before filtering for qualifying file operations.

All ten axes remain partial over the named inventory. External host session serialization, compaction summaries/checkpoints, initial content origins, some caller controls and external Markdown loading prevent complete classifications; known local witnesses do not resolve those alternatives. No observed or causal basis is supplied. Model parameters and static shipped guidance are not automatically part of accumulated target memory. The included opaque host continuation boundary is retained separately from the inspectable extension snapshot payload.

Source-read log: no target-source files were read for this classification. All warrants come from the supplied accepted records and their annotations; no records, quotations or new evidence were added.