Skip to content
Spectral Codex
GitHub
Decisions

ADR-079: Retire memory-tier signal path to Worlds

Context

ADR-018 redesigned Spectral’s memory system from five tiers to three (T1 Cycle Memory, T2 Run Memory, T3 Workspace Memory) and introduced a routing pattern at the T3 promotion boundary: T3 curation classified each observation as a domain observation or an agent-specific operational observation; domain observations were surfaced as structured world-signal events to spectral.worlds through the event path established in ADR-017.

Two of ADR-018’s load-bearing dependencies have retired under the in-band decision-support shift:

  • ADR-017 (event-driven signal path from spectral.platform to spectral.worlds) retired as part of ADR-074 (scan-pipeline retirement). The signal-routing infrastructure that ADR-018’s T3-into-worlds direction relied on no longer exists.
  • ADR-056 (T3 Memory → Worlds routing protocol) retired as part of ADR-074. The specific protocol shape for T3 memory events crossing the context boundary is gone.

ADR-018’s T3-into-worlds direction is therefore already structurally dead. This ADR makes the retirement explicit at the ADR-018 source so the source ADR’s content reflects the current architecture rather than leaving readers to discover the retirement by following supersession chains.

The principle ADR-018 D2 embodied — that long-term agent observations feed world-model evolution — does not migrate to a new construct. Under the shift, the platform→worlds signal stream is the world agent’s override-pattern detection on approved operator overrides at decision time (per ADR-081), which is a different mechanism with a different source. Agent T3 memories are no longer the source of evolution signal.

Decision

The T3-into-worlds direction and the T3 curation classification step retire from ADR-018. The three-tier memory architecture (5→3 tier redesign, T1/T2/T3 vocabulary) carries forward unchanged. The World Agent memory design (governed by ADR-058) carries forward unchanged.

What retires

  • T3 promotion boundary as a routing decision: the classification step at T3 (domain observation vs agent-specific operational observation) retires. There is no longer a domain-observation route out of T3, so the classification has no consumer.
  • World-signal events routed from T3 curation to spectral.worlds: the event substrate (ADR-044) and the contract-surface payload pattern (ADR-065) carry forward as substrate, but no T3-derived signal stream rides them.
  • ADR-017 dependency on the T3-into-worlds direction: ADR-017 itself retired in ADR-074; the dependency dissolves.

What carries forward

  • Three-tier memory architecture (T1 Cycle Memory, T2 Run Memory, T3 Workspace Memory). The 5→3 tier redesign decision and the elimination of T4/T5 stand.
  • T1/T2/T3 vocabulary as used by ADR-058 (World Agent memory). The tier names continue to identify scope and durability for agent memory storage; ADR-058 is the authoritative source for the World Agent memory’s per-tier semantics under the shift.
  • The “agents-have-memory, records-aren’t-memory” distinction captured in project memory: T3 is for an agent’s value-added observations about records, not for records themselves. This framing is unaffected by the retirement.

What does not migrate

The principle that long-term agent observations feed world-model evolution does not migrate to a new construct. The new platform→worlds signal stream is the world agent’s override-pattern detection on approved operator overrides at decision time (per ADR-081 + ADR-074’s principle-migration tracking for ADR-057). Source, shape, and consumer are all different. T3 memories under the shift exist for the owning agent’s own reasoning; they do not feed worlds for evolution.

Alternatives considered

Leave ADR-018 unchanged; rely on ADR-074’s retirement of ADR-017 and ADR-056 to communicate the retirement of the T3-into-worlds direction transitively. Rejected. A future reader landing on ADR-018 sees the T3-into-worlds direction as authoritative and must trace the dependency chain (ADR-018 → ADR-017 / ADR-056 → ADR-074 / ADR-079) to discover the retirement. The source ADR should reflect the current architecture.

Fully supersede ADR-018. Rejected. The three-tier memory architecture (T1/T2/T3) is in active use by ADR-058 and ADR-059 with unchanged vocabulary. Full supersession would force restating the tier redesign rationale without new content. Per discipline, this is a partial supersession of the T3-into-worlds direction only.

Migrate the principle (“agent T3 memories feed world-model evolution”) to a new mechanism inside the shift. Rejected. The shift’s evolution-signal mechanism (override-pattern detection on operator overrides per ADR-081) is structurally different — not T3-derived, not agent-memory-shaped, sourced from decision-time operator behavior rather than aggregated agent observations. Conflating them under a “T3 evolution direction migrated to X” framing would mislead future readers about the new mechanism’s actual source.

Consequences

  • ADR-018’s T3-into-Worlds direction segment (Decision paragraph 2 and the related Consequences bullets) is replaced with a one-line pointer to this ADR. The three-tier architecture decision (Decision paragraph 1) and the T4/T5 retirement consequence stand.
  • Codex updates flow: system-design/agents/memory-system.mdx rewrites to remove the T3-classification + world-signal-event path; the three-tier framing stays. ADR-058’s memory-tier vocabulary carries forward as the authoritative reference.
  • No Linear scope changes are introduced by this ADR beyond what ADR-074 already triggered (the scan-derived signal cluster closed there).
  • The override-pattern detection mechanism (the new platform→worlds signal stream, per ADR-081 and ADR-074’s principle-migration tracking for ADR-057) is the canonical signal source for world-model evolution under the shift. Its design specifics are deferred Phase 3 work (Spectral Agent reframe / World Agent role-expansion area).