Skip to content

ADR-0093: Keep an independently selected, evidenced memory model

  • Status: Accepted
  • Date: 2026-09-11
  • Related: ADR-0024, ADR-0057, ADR-0077, ADR-0086
  • User authorization: use GPT-6 Astra and Claude Fable 5.1 for chat while temporarily retaining GPT-5.6 Sol for memory formation

Context

Memory extraction inherited the interactive agent's model policy. Changing the chat default therefore changed the extraction model and invalidated its activation evidence. The Astra evaluations failed existing gates; inspection found plausible wording mismatches as well as omissions, duplicate memories, and classification differences. These results do not establish an intrinsic memory advantage for Sol. They do leave Sol as the evidenced deployment choice.

Decision

  1. The production chat default becomes astra; flagship and fable select Claude Fable 5.1. Retain balanced as an explicit GPT-5.6 Sol policy so the existing memory evidence retains its exact model-policy identity.
  2. formation.model_policy in memory/profiles.yaml independently selects the maintenance model through the existing static router. It defaults to balanced and is operator-overlayable. It is a policy-name reference included in the configuration inventory alongside the other maintenance controls. Invalid names fail profile validation; unresolved or unevidenced selections keep the existing auto fallback and required refusal.
  3. Non-routed fake compositions retain their deterministic model policy. Explicit memory evaluation modes use the requested evaluation model policy, allowing Astra to be evaluated without changing the deployed memory choice. off still performs no extraction-model resolution or provider call.
  4. Activation, extraction requests, cost accounting, and selection audits use the actual memory model policy. Audits also retain the chat policy. No evidence artifact, formation prompt, scorer, threshold, or trust gate changes.
  5. The existing content-addressed agent versioning preserves the immutable previous agent when the default model policy changes. New sessions use Astra; existing sessions retain their agent version and provider pins. Deployment continues to require draining runs whose registry versions are being replaced.
  6. Memory formation remains active Milestone 21 work. Revisit scorer semantics and compare Sol/Astra under controlled reasoning settings before attempting a memory-model change; activate a different tuple only after valid evidence.

Consequences

Chat upgrades no longer silently disable evaluated memory. The separate model selection remains visible in configuration and audit events. This is a fixed maintenance-model selection within the existing gateway, not authorization for dynamic routing, learned policies, or any other deferred roadmap item.