Every Ultron session is persisted. Transcripts are saved turn-by-turn as JSONL. Sessions can be resumed exactly where they left off.
Context injection at session start
Before you type your first message, Ultron loads context in a specific order:
Session start — context assembly
├── 1. Static system instructions — Ultron behavior rules, tool use guidelines, quality gates
├── 2. User profile — business name, ICP, voice tone, competitors, platforms
├── 3. Connected integrations — list of active services (Gmail, Stripe, HubSpot, etc.)
├── 4. Relevant memories — max 5 entries selected by Sonnet-powered recall
└── 5. Current date — injected as system reminder
Session persistence
- Transcripts saved as JSONL — one JSON object per message turn
- Includes: message content, tool calls, tool results, token counts, timestamps
- Resume at any time — context is fully reconstructed from the transcript
- Cost tracking accumulated per session and per-turn
Automatic compression
As conversations grow, the compression engine manages context automatically. You never need to manually summarize or restart a session:
- MicroCompact fires first — removes stale tool outputs from old turns (no API call)
- Session Memory Compact — uses stored memories as compression baseline (no API call)
- API Digest — full LLM summary via fast model (last resort)
After each compression, a boundary marker is inserted. All subsequent processing only touches messages after that boundary, preventing repeated compression of already-compressed content.
Cost tracking
| What’s tracked | Where to see it |
|---|
| Input tokens per turn | Settings → Usage |
| Output tokens per turn | Settings → Usage |
| Model used per call | Settings → Usage |
| Total session cost | Settings → Usage |
Long research sessions with many tool calls accumulate tokens quickly. Check usage after complex sessions to calibrate which tasks to run in API mode vs sandbox mode.
