Recall (parcha-skills) memory profile

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

Comparison rationale

Boundary of the compared memory. Recall ships only instructions, so the retained material it governs is mostly written by others. The host writes the transcripts, and the absent engine and Brain service write the index and remote records. The scope therefore includes every retained object that the shipped procedure reads or causes to be written. It excludes the skill's own static instructions and the current-run material Recall does not persist: search rows, windows, answers and receipts. Search rows are cited under authority and trace learning only as the delivery form of index content, not as retained objects. Most engine and service values carry basis claimed because the evidence for them is the skill's declared contract (RT-ABS-engine-code).

Write agency for user-started writes. Capture and skill-ify start on a user request, but the model selects or authors the content, and the procedure asks for no approval of that text. Under the profile type, starting a workflow does not make the admission manual, so both are automatic. The memory analyst described them as "user-triggered rather than automatic". That wording is about the trigger and does not contradict this classification. recall_forget is manual because the user's request picks out the exact record to withdraw.

Trace learning. For the same reason, capture and skill-ify count as automatic trace-fed writes: their durable outputs reach later sessions as evidence or instructions. Engine ingest also counts. The memory analyst's annotation on the index says its output is "retrievable evidence, not behavior-shaping instructions". The profile type does not require instruction force, though: a retained trace-derived artifact that a later consumer receives as context qualifies. The index's ranking and snippets reach later invocations in exactly that way, so the value is "yes" at basis claimed. The same route appears under behavioral_authority as a separate unit: the transcripts carry learning authority at the ingest update consumer, at basis claimed, apart from the index output's downstream ranking and routing. Skill-ify likewise has learning at its update consumer and instruction downstream. The skill-ify finding is afforded because its later-consumer leg runs through the host loader. Brain sync stays not-determinable because its admission control is unknown. The system union is "yes" with partial coverage. That supports no claim of improved capacity.

Curation. Capture and recipe separation compress retained session content into reduced standing records, so both are consolidate. The epistemic member calls their truth-apt relation indeterminate and finds no evidence of non-entailment, so synthesize is not supported. recall_forget is decay at basis claimed: the wired part is the agent's call, and the forgetting is the service's declared behavior. Its unit is partial because it would be invalidate if the service kept history, which is not stated. Host retention deletion of transcripts is a second decay unit, also claimed, because it is host behavior the skill describes. Generated-skill removal is recorded only as manual file deletion with no route, so it stays not-determinable under curation and write agency. Indexing and line redaction establish no controlled operation. Engine and service curation remain unresolved.

Withdrawal under write agency. Host retention deletion is automatic: the host deletes under a retention setting, with no per-content decision. recall_forget is manual, as above.

Read-back. Pull is classified only for routes that read session material, as reconciliation disposition 9 requires. Generated-skill direction depends on the host and stays unresolved. As a result, read_back_signal has no classified push unit, and its axis assessment is not-determinable, not inapplicable.

Supersessions that affect citations. Four supersessions change which IDs this profile cites:

None changes a value. Only the first two IDs carry classifications here.

Missing facts that prevent stronger assessments:

  • whether the engine uses dense representations, which limits form;
  • how synced and imported Brain content is encoded and admitted, which limits form, write agency, trace learning and trace source;
  • the external skill creator's admission behavior;
  • who removes generated skills, and by what decision;
  • whether captures draw on tool results;
  • the host loader's delivery direction;
  • service deletion semantics, which separate decay from invalidate;
  • whether index entries survive deletion of their source transcripts, which affects lifetime but no axis value.

No execution evidence exists, so no finding is observed.

Source reads. None. Every classification rests on the accepted records, which resolve every ambiguity this profile needed.