Enoch memory profile

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

Comparison rationale

This profile classifies accepted records, not additional source evidence. All positive findings concern wired implementation except the explicitly claimed learning-artifact descendant route. None asserts observed activation, faithful model interpretation, improved future capacity or causal benefit. Coverage remains partial on every axis because included native payloads and deployment-specific state are unresolved, not because source-only evidence cannot establish wired capacity.

The local JSON/JSONL/Markdown inventory supports files, not a database, repository or vector substrate merely from serialization or task repository use. Session ID mappings are access metadata; they do not reveal native storage or representational form. Static shipped instructions and external inference weights are not use-accumulated memory. See MEM-OBJ-codex-sessions, MEM-OBJ-native-context and RT-RTE-extension-loading.

Authority follows the actual consumer. Descriptive facts supply knowledge; confidence/time fields supply ranking; private identity supplies standing guidance. Trace inputs supply learning authority at automatic update consumers, while derived reasons and briefs supply knowledge or directions downstream. Stored checkpoints, statuses and reference agreement govern routing, eligibility and validation. Retained failed-check diagnostic prose supplies knowledge to a later repair model through bounded excerpts, separately from the doctor check's publication gate and review receipts; see RT-OBJ-validation-diagnostics and RT-RTE-task-work. Unconsumed provenance has no inferred authority. See MEM-RTE-startup-context, MEM-RTE-evidence-update, MEM-RTE-task-snapshot, MEM-RTE-candidate-curation and WF-RTE-review-handoff.

Admission is independent of authorship and I/O. Model-final fact saves and derived proposal/brief writes are automatic even when a human starts their workflows. The explicit human-requested branch is manual only at its own content-specific decision. Generic callers and human actor labels do not authenticate such a decision. Private identity installation, initial contents and provider checkpoint replacements remain unresolved. See MEM-RTE-memory-admission, MEM-OBJ-self and EPI-RTE-candidate-removal.

Curation follows transformations, not route names. Bounded summaries/extraction consolidate; normalized-text merges deduplicate and revise metadata; increased confidence can promote startup salience; physical forgetting removes facts; candidate retirement and later regression invalidate reliance while retaining history. A command named synthesis does not establish a new claim absent from inputs. Accepted epistemic records retain indeterminate semantic relations, so synthesize is unresolved rather than asserted. See MEM-RTE-memory-admission, MEM-RTE-task-snapshot, MEM-RTE-candidate-curation and EPI-RTE-evidence-synthesis.

Read-back chains separate a requested summary/display from automatic later supply. Startup is coarse availability/rank/budget selection, not identifier targeting merely because IDs appear in text. Activity-note flush selects a nonempty clipped operation summary coarsely; its conversation key identifies the receiving session, not which retained note parts to deliver, so that route carries no identifier signal. Actual session/checkpoint/candidate identity matches do select operative retained parts. Model selection of relevant prior context and candidate recommendation supplies inferred judgment. Native internal selectors remain opaque. See MEM-RTE-startup-context, MEM-RTE-task-snapshot, MEM-RTE-evidence-update and MEM-RTE-candidate-curation.

The trace-learning union retains yes from automatic retained task briefs and trace-fed proposals; activity and brainstorming branches keep their own provider/mixed-input limits. Conversation messages remain session-log sources, while task lifecycle and operation outcomes supply event sources. Neither a learning-artifact label nor raw-log retention closes a later-consumer route. No no finding or inapplicable trace-source unit is inferred from missing observation. See MEM-RTE-evidence-update, MEM-RTE-task-snapshot, MEM-RTE-activity-sync, MEM-OBJ-learning-artifact and RT-RTE-evolve-admission.

Reconciliation supersedes the duplicate epistemic EvidenceSignal, SemanticCuration and LearningArtifact objects with their memory declarations, and the epistemic published-skill route with the memory route. This profile cites those canonical destinations; the supersessions change identity, not evidence strength or semantic warrant. Valid containers and their distinct parts remain separate, including remembered text/metadata, candidate drafts versus durable candidates, and lineage assessment versus inbox. No supersession upgrades a classification or removes an opaque alternative.

Source-read log: none. No source ambiguity required a direct read; classifications use only the supplied accepted members and reconciliation. Missing native checkpoint contents/control, initial-content principals, promotion producers, scheduled-context refresh, lineage invalidation and installed plugin implementations prevent stronger coverage, without erasing the positive local mechanisms.