Self-Evolution: learning from past runs

Arbor remembers what it learns. Every run leaves concrete, situational findings — a dataset quirk that lifted the metric, a trap an executor or the harness fell into — and a later run on a similar task can start from that experience instead of from scratch.

There are two halves, and you only ever touch the second one.

Capture (automatic)

While a run is in progress, the Coordinator can log a finding the moment it hits one, via the RecordFinding tool — for example:

• leverage — "the dataset's labels above index 9000 are noisy; dropping them adds ~2 points"
• pitfall — "the executor kept editing eval.py; remind it the harness is protected"

At the end of the run, Arbor consolidates these (plus, optionally, findings it mines from the run itself) into an EXPERIENCE.md inside that run's session folder:

.arbor/sessions/<run_name>/
  EXPERIENCE.md      # the run's concrete findings (leverage / pitfall)
  findings.jsonl     # raw findings logged live during the run
  trajectory.jsonl   # decision trace (for SFT/RL)

This is on by default — it only writes a notes file, costs nothing extra, and is what makes the next run smarter. To turn it off:

research_config.yaml

coordinator:
  distill_skills: false        # don't write EXPERIENCE.md
  distill_abstract: false      # (opt-in) also LLM-mine the run for extra findings

distill_abstract is the only part that spends extra LLM calls — it asks the model to surface concrete findings the agent never logged explicitly. It's off by default; turn it on when you want richer experience and don't mind a couple of calls at finalize.

Reuse (you decide, in the intake conversation)

The next time you start a run in the same project, the intake conversation checks whether earlier runs left experience. If your new goal is similar, the planning agent offers it and asks whether to reuse it — you stay in control:

You: optimize the kNN solver for speed
Arbor: I found experience from a past run on this project
       (a GEMM + argpartition win, a cKDTree dead-end, a macOS taskset gotcha).
       Want me to start from those findings? [yes]
You: yes

When you agree, the relevant findings are composed into a short priors block and prepended to the instruction the Coordinator runs on — so it begins already knowing the dead-ends to skip and the directions that worked. Findings seen across several past runs are tagged [xN] and ranked first, so repeated lessons carry more weight.

Nothing is forced: the priors are presented as candidate directions to verify, not rules. Arbor's normal dev/test discipline still filters them — a stale or wrong prior just becomes a quickly-pruned branch, never a silent corruption of the result.

What experience is (and isn't)

Findings are kept specific on purpose. The value of "at d=16 spatial trees collapse; fuse distance + selection" is the detail — it's what lets the next run skip the same exploration. Arbor does not try to abstract findings into generic principles ("prefer partial selection over full sort"), which read well but rarely help.

Experience lives per session, not in a global library, so it never bloats the curated Skills menu. Recall is scoped to the current project, so an unrelated task never inherits another domain's tricks.

In short

	Who triggers it	Default
Capture (EXPERIENCE.md)	automatic, at finalize	on
LLM mining (distill_abstract)	config flag	off
Reuse at intake	you, in conversation	offered when a match exists