ADR 0002 — Memory consolidation boundary
ADR 0002 — Memory consolidation boundary
Section titled “ADR 0002 — Memory consolidation boundary”Status
Section titled “Status”Accepted.
Context
Section titled “Context”The runtime needs durable continuity without letting every turn write arbitrary long-term facts. The current live post-turn consolidation path is rule-based. LLM-assisted consolidation code exists as optional implementation work, but it is not exported or wired into the default post-turn path. Any future LLM-assisted path must preserve provenance, privacy, confidence, and operator correction flows.
Decision
Section titled “Decision”Memory consolidation is a post-turn job that writes through the Memory trait.
The active path uses rule-based extraction and persistence. If LLM-assisted
consolidation is reconnected later, provider timeout, provider failure, invalid
JSON, or empty output must degrade to the rule-based path rather than blocking
the turn. Persisted consolidation records use a stable slot key, semantic layer,
system provenance, and private default privacy.
Consequences
Section titled “Consequences”- The current rule-based path does not depend on provider calls.
- Any future provider/parser failure in the optional LLM-assisted path must not block the turn; it must downgrade to the rule-based path.
- Consolidated facts remain inspectable and forgettable like ordinary memory events.
- Tests should cover extraction/parsing, fallback behavior, and persistence through the memory trait rather than direct file or database writes.