SideNote ProManage AI conversations, history and context
Two different things are called *the conversation*: what is stored on your disk, and what is sent to the model. Understanding the difference explains most of the surprising behaviour people hit in long threads.
History and context are not the same thing
Local history
Everything ever said in the thread, on your disk, complete and searchable. Nothing is ever removed from it by compaction.
Model context
The subset actually sent with your next request, bounded by the context budget. Default 24,000 estimated tokens, adjustable from 4,000 to 500,000.
This is why a model can seem to forget something you can still scroll up and read. Your archive kept it; the request that produced the answer did not include it verbatim.
Context compaction, step by step
When a thread exceeds its context budget, SideNote Pro does not simply drop the oldest messages. It compacts.
- Keep the recent messages intactAt least the eight most recent are carried forward word for word, because recency is where most follow-up questions point.
- Summarize the older contextSideNote Pro asks your configured model to produce a structured summary of the earlier part of the conversation.
- Carry the summary forwardSubsequent requests include that summary in place of the original older messages.
- Save the boundaryThe summary and the compaction boundary are stored in the thread, so reopening it later behaves consistently.
- Keep every original message locallyCompaction changes what is sent. It never rewrites or deletes your local archive.
Compaction is a provider request. It uses your configured model, your key, and your tokens, and it adds latency when it happens. It is not free local summarisation, and on a high-reasoning profile it summarises at high effort too. This is a good reason not to leave a heavyweight profile selected for casual chatting.
Setting the context budget
The budget lives in Settings, History and context. The default of 24,000 estimated tokens is a deliberate middle: large enough for substantial documents and long threads, small enough that requests stay quick and affordable.
| Situation | Direction |
|---|---|
| Large documents rejected as oversized context | Raise it |
| Long analytical threads losing earlier detail | Raise it |
| Responses feel slow and costly on a hosted provider | Lower it |
| Running a local model on modest hardware | Lower it |
| Frequent compaction interrupting the flow | Raise it, or start new threads more often |
The estimate is heuristic, based on text length plus a fixed allowance per message and image, not an exact model tokenizer. It is close enough to budget with and deliberately not presented as precise.
Often the better fix is not a bigger budget but a new thread. A conversation that has wandered through four topics carries three of them as dead weight into every request. Ctrl+N is cheap.
Finding and organising threads
Ctrl+H opens history, grouped into Pinned, Today, Yesterday, Previous 7 days and Older.
- Search runs full-text across titles and message content, so a thread can be found by something said in the middle of it.
- Pin the threads you return to. Pinned threads are exempt from retention pruning.
- Rename when the generated title turned out not to describe the conversation.
- Delete, with confirmation.
- Export to Markdown for one thread.
- Open and continue, with your draft and scroll position restored.
Titles are generated from your first message or attachment name, shortened to about 52 characters. That makes the first sentence you type worth a moment's thought: it becomes the label you will scan for weeks later. Organising AI conversations has more on this.
Retention, protection and deletion
- History off. Disable local history entirely.
- Retention. Delete inactive threads after a number of days.
0keeps everything indefinitely. Pinned threads are never pruned. - Encryption. Apply Windows DPAPI to the archive, binding it to your current Windows account. Local at-rest protection, not a portable password-based format.
- Delete all history. Permanent removal of the archive.
A sensible default for most people: history on, retention at 90 days, pins for the threads worth keeping, encryption on if the machine is shared or portable.
Backing up
There is no cloud copy and no cross-device sync, so a backup is your responsibility. The archive lives under %LOCALAPPDATA%\SideNote Pro\.
One caveat that matters: a DPAPI-protected archive is bound to the Windows account that created it. Copying the file to another machine or another account will not make it readable there. If portability is what you want, Markdown export is the mechanism, not file copying.