Skip to content

Learning from traces

Most agent SDKs ship the runtime; few ship the improvement loop.

fastaiagent.learn reads completed traces out of local.db, extracts durable per-user / per-project / per-agent facts via an LLM, and re-injects them into future runs through PersistentFactBlock. The loop runs entirely on the developer's local machine — no platform dependency.

This is the SDK's take on the "continual learning" framing Harrison Chase has been writing about: traces are the substrate; agents improve along the context layer (memory, scoped facts, learned skills) without retraining.

Layout

Doc What it covers
Memory loop The end-to-end flow: traces → fastaiagent learn → learned_memory table → PersistentFactBlock
fastaiagent learn CLI Flags, scopes, dry-run, conflict resolution
Self-improving agents Conceptual framing — what we extract, what we don't, why memory-only at v1

What v1 ships

  • Memory only. Durable user/project/agent facts. No skill extraction, no prompt mutation. (Those need replay-eval to avoid drift — out of scope for v1.)
  • Offline batch. A fastaiagent learn CLI runs over the trace window you specify. No online mid-run learning.
  • Local first. Reads local.db, writes the new learned_memory table. Push to platform is unidirectional and unchanged.
  • PII-safe by default. Only --scope agent runs without an opt-in; user / project scopes require --allow-personal.

What's coming next

Tracked as future work in the plan file:

  • Skill extraction (reusable mini-procedures).
  • Meta-Harness style prompt/harness mutation.
  • Replay-eval infrastructure (prerequisite for both above).
  • Online mid-run learning.
  • UI for human review / annotation of learned facts.