Exocomp memory profile

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

Comparison rationale

This profile classifies accepted records, not fresh source inspection. No source files were read for this job. The members supply implementation and doctrine/design evidence, with no current-revision observed operation or causal experiment used. Every axis remains partial because the recorded inventory includes opaque provider retention and arbitrary program-created stores, plus narrower unresolved mechanisms. Local known units state complete classification only for their named operative part and path; they do not claim an exhaustive system union.

Page payloads are not uniformly natural language. EPI-OBJ-website-content distinguishes natural-language page text from symbolic HTML/Markdown serialization, and EPI-RTE-website-fetch carries the format-selected result into messages eligible for later retention. The form inventory separates these parts from conversation/answer prose and includes both alternatives. This classification neither authenticates page claims nor resolves extraction fidelity.

Recovery supplies accumulated messages automatically, while Await, Inquire, project List/Search and operator GET fulfill requests. A caller-selected name or path is therefore not itself an identifier-based push. Role discovery is different: code matches parsed Role to a template key before replacement. Restored checkpoint and message-bundle supply are coarse. Body selection at explicit skill Load is pull; automatic subsequent supply of retained injected bodies is coarse message delivery. See MEM-RTE-session-recovery, MEM-RTE-role-discovery and RT-RTE-skill-loading.

The Inquire summary is an afforded automatic trace-fed durable context artifact: recorded conversational prose feeds a summarizer, whose return can enter a Session tool message, backup and later inference. This supports trace-learning yes and session-logs at that local basis even though the original transcript remains and no improvement is demonstrated. Its update consumer gives retained prose learning authority; the downstream summary gives knowledge authority. Projection omits tool/system messages, so it does not establish tool-traces as a source for that qualifying write. Model condensation is consolidation, not evidenced synthesis; uncertain semantic fidelity neither removes consolidation nor proves new claims. See RT-RTE-agent-inquire, EPI-RTE-inquire-projection and MEM-OBJ-agent-messages. Raw checkpoint/snapshot copying alone does not qualify. No local trace-learning no finding is asserted without bounded absence evidence.

Content and control parts have different forces. Bug prose gives context; the fixed flag enforces tool-list exclusion without erasing JSON. Changelog descriptions and tuple metadata veto duplicates; dates affect relative order. Requirement status enforces omission from pending JSON, and source-first identity matching can discard authored Behavior. That forgetting supports decay rather than history-preserving invalidation. A reconstructed declaration is compiled from source, not extracted from an activity trace. Surviving Behavior supplies task directions and reference material, but the ordinary new Define* tool route is obstructed; direct definition replacement remains afforded. See MEM-RTE-project-records, MEM-OBJ-changelog-entry and MEM-RTE-requirement-update.

Local role registration affects later same-process lookup and does not establish which prompt/grants a fresh child executes. Skill catalog refresh, loaded-object adoption and system-body insertion are separate routes. Qualified Load/Unload changes loaded state and generic notices; only helpers or bare-name API use afford body insertion/removal. Removal can leave older differing bodies and never rolls back script effects. Cross-restart message-body retention cannot resolve loaded-state recovery. These distinctions prevent upgrading afforded instruction/extension routes to default wiring. See MEM-RTE-role-discovery, RT-RTE-skill-loading and RT-ABS-skill-grants.

Software admission is automatic even when a human starts or restarts the workflow. Explicit initial human supply/edit/approval would establish manual, but generic caller identity and arbitrary project authorship do not establish that decision. Known imports, recording paths and later replacements do not resolve initial provenance. Qualified human-answer messages have implemented advisory delivery and potential recovery; only the bare-name alternative reaches the bounds defect. This does not authenticate authorship or guarantee persistence, browser delivery, ordering or restored pending waits. See EPI-RTE-human-answers, MEM-RTE-session-recovery and MEM-RTE-project-records.

Reconciliation supersedes the epistemic bug-description and changelog-entry duplicates with MEM-OBJ-bug-description and MEM-OBJ-changelog-entry. Those canonical records support the corresponding substrate, form, lineage, authority and maintenance findings; supersession changes no value. Other overlapping route/container records remain distinct. The current pinned epistemic answer route agrees with the memory account; the memory report's older-fragment objection is not carried as a current conflict.

Missing records for provider-private payloads/updates and arbitrary program stores prevent positive service-object, parametric, learned-parameter, semantic-selector or additional trace-source classifications. Missing initial authorship/control, live extension-state reload and automatic diagnostic consumers prevent complete lineage, agency, read-back and authority assessments. These are limitations, not evidence of absence. Static role templates, model components and binary installation are not substituted for accumulated memory merely because they shape the runtime.