oh-my-pi memory profile

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

Comparison rationale

The accepted inventory is a partial loop. Every axis therefore keeps its inspected positives beside unresolved included parts. A wired finding establishes a conditional implementation path, not observed admission, delivery, compliance or benefit. No bounded absence record was declared, so this profile asserts no trace-learning negative or exhaustive absence of other controlled values.

The managed-skill object supersession affects storage, form, lineage and downstream authority: the canonical object is EPI-OBJ-managed-skill, with EPI-OBJ-managed-description separately owning applicability text. The superseded memory object's findings remain part of the accepted account, but citations here use canonical objects and the unsuperseded MEM-RTE-autolearn-capture for trace-conditioned writes. No valid local-memory, Mnemopi or decision-file container is superseded by its parts. Grouped units above name the operative parts, not a uniform classification inherited from those containers.

Local memory summaries/lessons provide contextual advice subordinate to current evidence. Sharpshooter decisions and stable mental-model text supply standing instructions. Retained experience at a consolidation/capture consumer has learning authority independently of the output's knowledge/instruction authority. These mappings follow MEM-BAP-memory-guidance, MEM-RTE-local-consolidation, MEM-RTE-decision-memory and MEM-RTE-autolearn-capture; they make no claim of improved future capacity. Opaque provider replay cannot inherit the readable compaction summary's form or force. Numerical columns do not inherit the surrounding prose's authority or prove a vector retrieval selector.

Explicit requested recall, file/row/history access and skill invocation are pull operations. Automatic context reconstruction and backend instruction supply are push operations. A leaf/parent/checkpoint match actually selects reconstructed history; a requested memory URL alone is not identifier-based push. First-turn relevance selectors remain unresolved even though query composition, bank filters and budgets are known in MEM-RTE-backend-recall. The profile does not turn FTS schema presence into inspected lexical ranking or an embedding column into an embedding selector.

Automatic extraction, consolidation and private capture admit transformed content without an independent per-content human decision. Explicit writes also have a separate conditional human-approval path in RT-RTE-tool-dispatch. The human approves that invocation's effective content-bearing parameters; the gate supplies neither factual acceptance nor evidence that a particular deployment exercised it. Generic initial/caller provenance remains unresolved rather than inheriting either control classification.

Trace learning is supported for local derived memory, decision replacements and private capture to durable lessons/skills. Raw journal retention and a service retain acknowledgement alone do not establish that transformation. Generated continuation summaries would qualify if their automatic trace-fed production is established; the accepted records expose method choice and later reconstruction but not complete producer algorithms across text, provider and image alternatives. Those parts stay unresolved rather than becoming a local no. Session turns remain session-log origins even when submitted through an event callback, and imported project documentation adds imported lineage rather than an event-stream source. Tool traces are positively supported in copied capture history, not inferred for every extraction window.

Curation findings distinguish deletion from history-preserving withdrawal. Local no-output deletes the older derived row; the retained raw source journal does not preserve that row's history. Managed directory deletion likewise establishes forgetting, not invalidation. Notebook empty replacement retains prior journal revisions, while Mnemopi retirement retains eligible rows with retirement metadata. Each transformation now has its own unit and citations: MEM-RTE-local-consolidation, MEM-RTE-notebook-update, EPI-RTE-skill-mutation and MEM-RTE-memory-edit. No unit borrows another mechanism's invalidation warrant; unresolved core propagation remains separate.

No source file was read for this profile. Classification used only the supplied accepted records and reconciliation. Missing core/vector provenance, checkpoint production, upstream mental-model seeding, numerical ranking, alternate API callers and complete invocation/host consumer routes prevent stronger coverage assessments. WF-RTE-result-delivery also limits optional child artifacts: a pointer/current return is not proof of durable subsequent retrieval. Static discovery instructions and ordinary current-run scheduling/format checks are outside the accumulated-memory boundary.