Stored with the account
The conversation — live messages, archived stretches and the checkpoints inside them — is one record per account on the server. It loads when you open the editor and saves 600 ms after every change (up to 600 saves an hour, one blob of at most 400 KB). Sign in from another browser and the same chat is there. Delete your account and it is deleted with it. There is one conversation per account, across all your documents; the panel header always says which document you are working on, and every proposal remembers which document it belongs to.
A live window and an archive
Each request carries the recent conversation verbatim. When that window grows past 12,000 tokens (about four characters each) and there are more than eight messages, everything older than the newest eight folds into an archive and only a summary stays in play. The summary is written by the fast tier — Claude Haiku 4.5 on Runic’s key — in up to about 150 words, with one rule that matters: every concrete value stays verbatim — step names, roles, amounts, thresholds, SLAs, retry counts, system and document names, each suggestion or decision as its own bullet — because “a threshold was agreed” loses the threshold. If that call fails or has not landed after 20 seconds, a deterministic note takes its place (“12 earlier messages, starting with: …”). Archiving never waits on AI.
In the panel an archive shows as “📦 12 archived” followed by its summary. Click the row (▸ / ▾) to expand the original messages; applied proposals inside still offer ↩ Restore, unapplied ones become stale. Up to 20 archives are kept.
Recall when you refer back
Summaries alone would lose detail, so Runic also recalls. Your new message is broken into tokens — words of four letters or more, stemmed to five characters, Turkish-aware, plus numbers — and compared with every archived message; any archived message sharing at least two tokens scores, and the best three come back verbatim under “RECALLED MESSAGES”, up to 4,000 characters each. The model is told to treat both summaries and recalled messages as things that really happened in this conversation — and, when a detail exists only in a summary and was not recalled, to ask for it rather than invent it.
Recall is plain word overlap, deliberately: no embeddings, no hidden state. If you refer to “the escalation step”, the message that mentioned escalation is the one that comes back.
What travels with each request
The live window — recent messages, each cut at 6,000 characters on a line boundary with a “[… truncated]” marker, 48,000 characters in total with the oldest dropped first; the server keeps at most the last 30 turns of 8,000 characters each. Up to six archive summaries of 2,000 characters; up to three recalled messages of 4,000; the names of your other documents; and the whole current document, up to 60,000 characters. Proposal bodies are never repeated in history — a message that carried a diff is sent as a bracket note like “[proposed an edit to ‘Onboarding’ — applied]”, because the document itself already travels with the request. A message that was truncated is a message the agent will ask about rather than guess from.
Limits
“Clear” removes everything, including archives and checkpoints. The saved record is capped at 400 KB (“Chat state too large.”); archiving drops the text payloads of unapplied proposals to stay well under it. The summary route allows 30 archives an hour and needs only a signed-in account, not a verified email; every failure falls back to the deterministic note. Memory is not per document: switching documents does not switch conversations. Older builds kept the chat in the browser; on first load Runic migrates that copy to the account and removes the local one.
How to use it
- Chat as usualNothing to set up. Once the conversation outgrows its window, a “📦 N archived” row appears above the live messages with a summary.
- Expand an archive when you need itClick the row (▸ / ▾) to read the original messages or to reach an older ↩ Restore.
- Refer back in your own wordsUse the words you used before — “the retry loop we discussed for payments” — and the matching archived messages are recalled for that request.
- Ask it to act on earlier suggestions“Apply what you suggested” builds the edit from the conversation instead of asking you to restate it.
- Pick up on another deviceSign in anywhere; the same conversation and checkpoints load with your account.
- Start over“Clear” in the panel header wipes the conversation and its archives.
Commands & shortcuts
| 📦 N archived | an archived stretch; click to expand or collapse |
| ↩ Restore (inside an archive) | checkpoints stay usable after archiving |
| Clear | wipe live messages, archives and checkpoints |
| [… truncated] | the marker on a history message cut at 6,000 characters — the agent asks rather than guesses past it |
Tips
- Name things consistently in chat. Recall is word overlap, so the noun you used the first time is the noun that brings the message back.
- You do not have to re-paste a suggestion: “do that” and “update it with what you suggested” are understood.
- If the agent replies that it needs a detail, that detail sat in an archive summary and was not recalled; give it in one line and continue.
- The header says which document you are working on — check it before applying a proposal after switching documents.
Questions
Is the conversation per document?
No, it is one conversation per account. Each proposal remembers its own document, and Apply only works on the document it was made for.
Where is the chat stored?
In Runic’s database, one record per user, saved shortly after every change. It follows your account to any browser and is deleted with the account.
Can I change when archiving kicks in?
The live window budget is 12,000 tokens. There is no settings control for it; the client reads an optional chatArchiveTokens value (minimum 500) from the runic.prefs entry in the browser’s local storage if you set one.
Does the agent see archived messages?
It sees their summaries on every request, and the original messages when your request shares words with them. It is told never to invent details a summary does not contain.