Conversation memory

Follow-up questions work because the model sees the thread — the last answer whole, every older answer as a one- or two-sentence summary.

Conversation memory ​

A follow-up like and for a signature mismatch? only makes sense next to what came before. DocPilot sends the model the thread, compressed: the last answer whole, and each older answer as its summary. The model writes that summary itself, in the same call that produces the answer, so memory costs no extra request.

What the reader sees
How do webhooks retry?
Failed deliveries are retried with exponential backoff: 1 min, 5 min, 30 min, then hourly, up to 8 attempts over 24 hours. Each attempt carries the same event id, so a handler must be idempotent…
Can I replay one by hand?
Yes. Open Webhooks → Deliveries, pick the event and press Replay. A replay is a new attempt with the original payload and the same event id, and it does not reset the retry schedule.
→
What the model is sent
userHow do webhooks retry?
assistantFailed deliveries are retried with exponential backoff: 1 min, 5 min, 30 min, then hourly, up to 8 attempts over 24 hours. Each attempt carries the same event id, so a handler must be idempotent…summaryRetries use exponential backoff, 8 attempts over 24 h, same event id.
userCan I replay one by hand?
assistantYes. Open Webhooks → Deliveries, pick the event and press Replay. A replay is a new attempt with the original payload and the same event id, and it does not reset the retry schedule.

How it works ​

  1. Every answer is returned through the answer tool, which has a required summary field: one or two sentences.
  2. The summary is stored with the turn (as answerSummary) but never shown to the reader.
  3. On the next question, the newest question-and-answer pair goes to the model in full, up to 2,000 characters.
  4. Every pair before it goes as the question plus its summary.

The model therefore keeps the details of what it just said, and the gist of everything earlier, without the prompt growing by a full answer every turn.

Configuration ​

There is nothing to set: this is how 1.5 builds every request. Two environment variables restore the earlier behaviour, for measuring the difference:

bash
DOCPILOT_HISTORY_MODE=skeleton   # older answers dropped, as in 1.4
DOCPILOT_ANSWER_SUMMARY=0        # no summary field in the answer tool

Measured against 1.4 on this site's eval set: clarifying follow-ups scored higher (+0.035 answer F1), rewrite requests slightly lower (−0.039), at about 6% more requests. The numbers and the arms that lost are in the changelog.