v2.12.0

An opt-in pipeline that mines a model's thinking from ~/.claude transcripts, distills it into reusable reasoning-pattern fibers, and injects the learned strategies into other models' sessions. Off by default; traces are redacted before they are stored.

Added

• reasoning_traces staging store (SQLite/SurrealDB/in-memory/base) + schema migration v39 → v40 + delete_reasoning_traces_by_model
• reasoning_miner — scans ~/.claude thinking, redacts secrets via safety/sensitive before insert (PROCESS_REASONING_TRACES)
• reasoning_distiller — segments → classifies → clusters into ReasoningBank pattern fibers, restricted to mining_models (LEARN_REASONING)
• reasoning_injection — SessionStart + UserPromptSubmit hooks inject learned patterns for the active model
• Dashboard reasoning page (/ui), smem reasoning CLI, smem_reasoning MCP tool (58 tools)
• [reasoning_training] config — mining_enabled / injection_enabled (both default off), mining_models, backfill + privacy controls

Fixed

• Hook output contract — hooks now emit {"hookSpecificOutput":{"hookEventName","additionalContext"}}; the old {"type":"context"} shape was silently ignored by Claude Code, so SessionStart memory/reasoning injection never reached the model
• Shared-storage brain race — mining runs on an isolated, non-cached storage instance (create_isolated_storage) so a concurrent request's set_brain cannot misroute its graph writes

Notes

• Fully opt-in and privacy-preserving: thinking redacted before storage, no secrets reach evidence, mining + injection default off

Full changelog: https://github.com/acidkill/surreal-memory/blob/main/CHANGELOG.md#2120--reasoning-training-learn-how-models-think