Reasoning saved across reloads

The "Thought for Xs" block from a reasoning model survives closing and reopening the chat.

Short answer

The “Thought for Xs” block from a reasoning model survives closing and reopening the chat. It reappears collapsed, click to expand and read the reasoning again.

Detailed explanation

When a reasoning model finishes generating, Abliteration stores three things:

  1. The final answer text (as normal)
  2. The <think>...</think> content (in a separate database column)
  3. The elapsed reasoning time (stored in milliseconds in the database, displayed in seconds in the UI)

On chat reload, the message is rebuilt with all three: final answer visible, collapsed reasoning card above it showing “Thought for Xs”.

Why this matters technically. Earlier generations of local model clients often threw away reasoning after generation - you would see the “Thinking…” card during the chat, but if you closed the chat, on reopen only the final answer would be there. The reasoning was lost. Abliteration keeps it.

Legacy chat fallback. If you have older chats where reasoning was saved as raw <think>...</think> inside the message content itself (not in the separate column), Abliteration detects and re-parses those on load, showing them as a collapsed card without the duration counter. This means old reasoning chats aren’t lost even if the storage format has changed.

Why it matters

Reasoning is often the most valuable part of what a model produces on hard problems - it’s the audit trail for how it got to the answer. Losing it on reload means losing the ability to review or cite the model’s process later. Persistence means reasoning is treated as first-class content, not throwaway.

In Abliteration specifically

Stored in the SQLite database alongside message content. Regular (non-reasoning) chats behave exactly as before - no schema surprises, no performance cost.

The reasoning card looks identical whether streamed live or restored from storage - same header, same expand/collapse, same copy button. If duration was recorded (new chats), header says “Thought for 12s”; if only the reasoning text is available (legacy chats), header says just “Thought” without duration.

Catalog is the map. Apps are the tools.

Run models on your own machine, not in the cloud.

Every model page has an "Open in Abliteration" button that hands the model directly to the first-party desktop client, at the quantization your rig can actually run. No API keys, no subscription, no prompt leakage.

Open in Abliteration