Recall is a third independent brief at pre-dispatch — fanned beside assemble and delta, never fused with either

The recall-push (the salient subgraph of the Cortex, rooted at space-0) leaves the assemble briefing and becomes a third independent brief, produced by its own subagent fanned at pre-dispatch alongside assemble (curated self-state) and delta (world-new). Three views, three subjects, three faithful agents (ADR-0004). The wake — and every dispatch's pre-dispatch — fans three briefs.

Status

proposed (2026-06-10; Voz ratifies, glossary grill). Amends ADR-0009 (briefing composition: the recall leg leaves the briefing and becomes a peer brief) and the recall-push placement inside assemble (feat(assemble): recall-push, 4428c64). Honors ADR-0004 (one task per subagent), ADR-0011 (navigation is judgment in the loop — reaffirmed; only the push is fanned), ADR-0012 (pre-dispatch widens from assemble + delta to assemble + delta + recall).

Considered options

• Fuse recall with delta in one shared-context subagent. Rejected on the subject boundary —

the glossary's strongest line: delta is over the world, never over the wiki; recall is navigation of own knowledge, never re-ingested. One context holding both is where the self-reference guard fails in practice: the edge's own recalled Artefato sits beside fresh world signal with only model discipline keeping one from being read as the other (the Zep-failure shape, transplanted to sources). The two reads also need different judgment (novelty-worth- surfacing vs salience-to-now) and different tools (source keys vs graph traversal).

• Keep the recall-push inside assemble (the status quo, 4428c64). Rejected on context budget

and fidelity: assemble's brief is already the fattest surface in the system (~60k tokens), and ADR-0011 already ruled that assemble is mechanical while Cortex work is its own thing. The push was placed there as a compromise; a dedicated agent can navigate deep (traversal + semantic search over past Artefatos) and return only the salient subgraph.

• Chosen: independent recall subagent. Parallelism is real — recall reads the durable graph

(space-0, Direction anchors, curated clusters, published Artefatos) and does not need the sweep's seconds-fresh output: what the sweep just extracted is non-curated, exactly the tier that is not yet salience-bearing. Cost: one-beat staleness on brand-new extractions — acceptable by construction.

Decision

• Recall becomes a noun — the yield of recalling (the act stays a lowercase verb), symmetric

to Delta (the yield of updating). The brief it yields is the memory-salient view: the salient subgraph of the Cortex, rooted at space-0.

• Pre-dispatch fans three briefs: assemble (curated self-state) · delta (world-new) · recall

(memory-salient). All three are read-only subagents; the loop blocks on assemble, the delta never gates (ADR-0011), recall degrades to briefing-only waking on a graph outage.

• The push seeds; navigation deepens. The recall brief is the push (mechanical salience,

fanable). On-demand Cortex navigation remains the loop's own judgment in its own window (ADR-0011's line, reaffirmed — it was never assemble's, and it is not the recall agent's either).

Consequences

• The skills now lag the decision (a build follows, not in this ADR): skills/wake ("fan the two

briefs" → three), skills/_shared/pipeline.md (pre-dispatch = assemble + delta + recall), skills/assemble + tools/briefing.py (the recall leg moves out), a new skills/recall subagent.

• The briefing (ADR-0009) returns to its four parts; the salient subgraph is no longer §7 of the

briefing but a peer brief beside it.

• If the Cortex grows to where the salience traversal wants its own cadence or budget, the

independent agent is already the right seam — that future tuning needs no further architecture.