Agents that improve themselves
Run 50 beats run 1 — reflection turns experience into procedure.
Without memory
Most production agents are exactly as good on their 500th run as their first — every mistake is repeated forever because nothing closes the loop between outcome and behavior. “Agent improvement” usually means a human editing the prompt.
With Rekall
After each run, a reflection pass mines what worked and what didn’t into procedural memory. Successful strategies become preferred workflows; failure patterns become things the agent checks for. The improvement loop is part of the runtime.
How it works
Runs end with reflection
The reflection worker reviews the run: outcome, cost, wrong turns, and what the human corrected.
Lessons become procedure
Recurring wins codify into procedural memory and skills; recurring failures become pre-flight checks.
The next run starts smarter
Run 51 recalls the codified strategy from runs 1–50 — improvement is compounding and inspectable, not vibes.
See it in action
[reflection] run #50 (data-backfill) completeoutcome: success, 22 min (median: 41 min)what differed: batched updates BEFORE indexrebuild (runs 12, 31 failed the other order)[procedural] strategy reinforced: batch-then-indexconfidence 0.91 (was 0.72)[skills] backfill.skill.md updated — step 3 nowmandatory pre-check
Related use cases
Skill codification — workflows that teach themselves
Work out a tricky procedure once. Rekall turns it into a skill your whole team runs.
Sleep-time compute
Memory that organizes itself while your agents are idle.
Scheduled agents that remember
A daily digest agent that knows what it already told you.