ADR-0082: Learn before you forget (mine the compaction-archived span)

Source: section ### ADR-0082: Learn before you forget (mine the compaction-archived span) of the project decisions log (project-iris-prd/13-decisions-and-rationale.md), extracted on 2026-10-18 for the Evarness repository. It predates the rename from IRIS and is kept as written.

Decision: When conversation compaction (ADR-0079) archives a span of old turns, capture that span so the behavior miner (ADR-0080, digital-twin L1) mines it specifically, rather than relying solely on the daily tick's 80-turn recency window. CompactedHistory now exposes archived_turns; _record_turn buffers them into a bounded in-memory _compaction_archive_buffer (cheap, no LLM, only when the miner is enabled); and _run_behavior_mining drains the buffer and mines it as its own pass before dedup + propose. A separate pass is required because render_activity keeps only the tail of the transcript (text[-max_chars:]), so folding the older archived span into the recent window would truncate it away. The mining stays on the heartbeat — _record_turn only buffers — so the hot chat path takes no LLM cost.

Rationale: The behavior miner reads the last ~80 turns. In a long or very active session, the oldest turns — often the richest for spotting a habit — scroll out of that window before the daily tick runs, and are never mined. Compaction is the exact moment the harness decides those turns no longer fit the working context; it is therefore the natural trigger to extract durable learning from them before they leave. This closes the loop between the two big context-management features: the same event that evicts a span for forgetting now also feeds it to the part of the brain that learns. The turns themselves are never lost (they persist in the conversation store + semantic index); what this fixes is the miner's blindness to spans that have aged past its window. Bounded buffer, propose-only, off unless the miner is on — no new always-on cost. Full design: docs/architecture/digital-twin.md (L1 note) + docs/architecture/memory-subsystem.md (compaction note); narrative docs/blog/2026-06-28-learning-from-what-you-forget.md.

---