Exo memory profile

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

Comparison rationale

All values are classifications of accepted records, not new source findings. The inventories follow the memory report's inspected and unresolved routes. No axis is complete: local witnesses do not resolve native payloads, alternate harnesses, host-specific selection or generic caller control. All positive implementation findings remain wired unless explicitly marked afforded; there is no supplied observed operation or causal comparison.

Fact and plan text, symbolic status/membership metadata, executable modules and opaque checkpoint contents are different operative parts. File storage does not establish textual form, and a readable ID/preview does not classify its payload. The local-profile branch includes mutable retained text, not the shipped static identity prompt. Repository source changes enter the inventory only through the recorded later rebuild/load path, not as a claim that all repository files are memory. See RT-BAP-harness-instructions, RT-RTE-guardian-update, MEM-OBJ-sandbox-snapshots and MEM-RTE-codex-compaction.

The fact store's authoritative-context wording gives instructional force, not verified truth. Skill membership supplies a local read veto, but direct supporting-file access survives unlisting. Retained approval correspondence is a criterion and veto on Basic recovery, not an oracle for answer quality. These consumers explain why instruction, knowledge, routing, validation and enforcement coexist without implying a universal memory policy. See MEM-RTE-fact-revision, MEM-RTE-skill-lifecycle and RT-RTE-managed-turn.

Trace learning here has the profile contract's automatic-write meaning, not the stronger claim of improved future capacity. Tool-result reduction, scheduler continuation formatting and native replay compaction have trace-fed writes and later consumers. Native compaction's opaque algorithm prevents semantic curation/form classification, not the recorded continuation chain. The update consumer receives learning authority separately from the later model's knowledge/context consumption. No presumed reflective criticism, novel knowledge or improvement is needed for these wired classifications. Foreground fact/skill/code proposals remain unresolved as trace learning: software admission alone does not identify their original input. See MEM-RTE-result-offload, MEM-RTE-codex-compaction and RT-RTE-background-wakeup.

Prefix reduction consolidates an existing retained result into less context; raw offloading or structural validation alone does not. Fact forgetting invalidates current reliance with version history retained, while capacity cap eviction forgets old current entries. Skill unlisting is only a bounded invalidation. No compaction command name establishes synthesis or a semantic summary. See MEM-RTE-result-offload, MEM-RTE-fact-revision, MEM-RTE-skill-lifecycle and MEM-RTE-codex-compaction.

Read direction is classified per operation: automatic catalog/context supply is push; requested skill bodies, event pages and checkpoint restoration are pull. Ordinary exact-path artifact loading does not establish identifier-based push targeting. Scheduler owner/task matching does select a retained result for its destination, unlike coarse availability/status/budget filtering. Tool and session inputs remain tool-traces/session-logs when serialized as journal events. An exported trajectory is not a qualifying learning input without a later automatic update consumer. See MEM-RTE-skill-lifecycle, RT-RTE-background-wakeup, MEM-RTE-codex-compaction and WF-OBJ-trajectory.

Reconciliation supersedes the epistemic memory-entry and todo object identities with MEM-OBJ-exo-memory and MEM-OBJ-exo-todos, and the epistemic todo-state route with MEM-RTE-plan-replacement. The profile uses those canonical destinations. These supersessions affect identity resolution, not a positive value or evidence strength. The combined epistemic skill object remains a valid container; separate memory catalog and payload records provide the units. Specialized write/read/forget records remain valid within the fact-revision lifecycle. No conflict is resolved by selecting a stronger status.

Source-read log: none. This classification used only the supplied members, types and contracts; no source ambiguity required a checkout read. No analyst record was declared, amended or newly evidenced.