Pi Posthorse memory profile

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

Comparison rationale

The memory boundary is accumulated task content and its operative access metadata, not every stored runtime object. Notes and session files supply the retained substrates. Git common-root resolution does not make notes repository memory. Transient projections, callback maps and page reservations do not add a durable in-memory substrate. Display metadata is not a second model-facing payload. These distinctions follow MEM-OBJ-notes, MEM-OBJ-session-journal, MEM-OBJ-posthorse-display and RT-CMP-page-reservations.

Automatic recovery is consolidation and trace learning under this comparison contract. It compresses retained conversation and tool receipts into a durable continuation that a later model receives. Neither novel claims nor measured improvement are needed for the wired trace-learning finding. Learning authority belongs to the traces at the update consumer; knowledge and fixed-warning instruction belong to the resulting checkpoint at the later model. The epistemic reshaping account does not contradict that classification or establish semantic fidelity. See MEM-OBJ-recovery-record, EPI-RTE-recovery-reshape and MEM-BAP-recalled-content.

Requested note/history returns are pull. Checkpoint projection and automatic reminder-metadata lookup are push operations. Literal query/path/entry targeting in requested recall is not inferred-lexical or identifier push. Automatic recovery combines coarse role/recency/budget selection with actual projection/provenance identity matches. Explicit-reset callback identity authenticates admission of a proposed write; it is not targeted retrieval of an older memory object. See MEM-RTE-notes-recall, MEM-RTE-history-recall, RT-RTE-explicit-reset, EPI-RTE-recovery-provenance and RT-RTE-request-guidance.

The positive automatic write findings cover model/software admission without a required per-content human decision. They do not classify generic callers, pre-existing notes or external editors as autonomous. Optional human authorship is not an evidenced manual admission mechanism. Native/manual checkpoints remain included: their later textual consumption and ancestry selection are recorded, but production control, original inputs and generation fidelity are not. See RT-RTE-note-admission, MEM-OBJ-notes and MEM-OBJ-recovery-record.

No supersession affects a value. Reconciliation retains valid containers alongside their parts and identifies no cross-report conflict. The profile therefore cites the original canonical declarations. Unresolved units remain local to each axis rather than being removed to obtain complete coverage. In particular, arbitrary note content and actual image payloads prevent complete form classification; imported note provenance and native checkpoint production prevent complete lineage/admission/trace classification. No unsupported absence or inapplicability is substituted for those gaps.

All positive findings concern implementation wiring unless their own basis says otherwise. Host persistence and delivery are external contracts; no supplied run or experiment establishes activation, improved capacity, reliable restoration or task truth. The selected workflow's scripted retrieval and reopening assertions do not upgrade the findings to observed operation: see EPI-RTE-reset-evaluator and RT-CLM-context-continuation.

Source-read log: none. This classification used only the supplied accepted members and contracts; no source ambiguity required reading frozen source files.