Thesis

V4's reasoning_content is a high-fidelity log of the model's internal decomposition — why it took an action. Today the engine sanitizes/placeholders it just to avoid HTTP 400 (the API requires reasoning_content replay on tool-call rounds). Inversion: capture every block as a reflective memory record (#536's taxonomy) tied to the producing turn. The HTTP 400 workaround becomes the load-bearing connection between the model's actual thinking and the persistent memory store.

Soft-blocked on #534/#536 landing.

Current behavior

• crates/tui/src/client/chat.rs:458-467 substitutes a placeholder string when reasoning_content is missing, just to keep the conversation chain valid for V4 thinking mode.
• crates/tui/src/client/chat.rs:688-714 scans every assistant message and forces a non-empty reasoning_content for replay.
• The actual reasoning blocks are surfaced to the UI as ephemeral content; not persisted, not queryable, not searchable.

Proposed change

• Intercept every reasoning_content block in the streaming response.
• Write it via MemoryBackend::remember (MEMORY: substrate — MemoryBackend trait + SQLite-backed store (FTS5) #535) as a reflective memory (MEMORY: engine — typed memory model + multi-signal recall #536), tagged with: turn ID, model, reasoning_effort level, file/symbol references extracted from the content.
• Tie to the resulting tool calls / file edits (graph edge causes per MEMORY: topology — graph edges + contradiction detection #537) so future "why was this code written this way?" can traverse from the artifact back to the thought.
• Recall from inside the workshop (Roadmap: finish workspace extraction and remove runtime integration seams #8) so the model can query its own past reasoning programmatically.

Open questions / risks

• Storage growth: every turn produces a reasoning block on Think-High/Max. Need pruning policy under MEMORY: hygiene — decay scoring, periodic review subagent, one-key drop #539 hygiene — probably "keep verbatim for N days, then summarize, then drop after M days unless promoted."
• Sensitive content: reasoning may include secrets/API keys/internal identifiers the user didn't intend to persist. Need a redaction pass on ingest, or per-scope opt-out.
• Cost: reasoning blocks can be long; ingestion subagent (MEMORY: loop — lifecycle hooks + background auto-extraction #538) needs to handle them efficiently — possibly via the workshop (Roadmap: finish workspace extraction and remove runtime integration seams #8) for the long ones.

Acceptance signals

• reasoning_content blocks are persisted via MemoryBackend.
• Recall surfaces the original reasoning when querying about a specific edit/file.
• Edge graph shows causes links from reasoning → tool call → file change.
• A turn 5 sessions ago can be queried: "what was the model thinking when it wrote function X?"