Abstract
Agents produce high-fidelity traces of everything they do. But traces alone don't produce memory. This talk introduces three primitives for turning session history into something actionable:
- Dreams: Structured interpretation over session traces. Not a model inventing lessons, but labeled observations anchored back to the evidence that produced them. Every observation carries a receipt: the session, turn, and outcome that caused it to exist.
- Reflections: Per-session summaries that compress what mattered without discarding the path. They make the trace searchable and comparable across runs without requiring a model to re-read thousands of tokens.
- Inceptions: Replay. Take a prior session, fix the task, run it with the dream's lesson applied. If the outcome holds across models and harnesses, the lesson belongs to the workflow, not to one lucky run.
The result is a feedback loop with evidence at every layer: capture the trace, label what mattered, find patterns across sessions, replay to confirm what's actionable. Skills your best engineers already have, captured, reflected on, and made replayable by the whole team.
Built on tapes, Paper Compute's open-source agent telemetry layer.
Date
Wednesday Nov 18 / 02:45PM PST ( 50 minutes )
Location
Ballroom BC
