Skip to content

EXPERIENCE.md

Latest commit

 

History

History
51 lines (40 loc) · 2.58 KB

EXPERIENCE.md

File metadata and controls

51 lines (40 loc) · 2.58 KB
Raw
tags
core
experience
learning
runtime
project ABSOLUTE_ZERO
status active
confidence high
date 2026-07-11
summary Experience Learning Engine spec — extracts lessons, failures, workflows, reusable code and patterns after every completed task.

ABSOLUTE ZERO — Experience Learning Engine

Learns from every completed task, automatically, from artifacts the OS already produces. Implementation: scripts/experience.py (stdlib).

1. API

python scripts/experience.py harvest [--trace <file>]   all unharvested closed traces
python scripts/experience.py recall "<query>"           semantic retrieval
python scripts/experience.py --selftest

Idempotent: harvested traces are marked "harvested": true and skipped.

2. Extraction → storage

extraction source stored where
Lessons a VERIFY-fail note followed by its retry EXECUTE note is a symptom→fix pair drafted into 30_LESSONS/<slug>.md (status: draft, confidence: low, trace cited) — real indexed notes; /review promotes or prunes. Deduped ≥0.5 sim vs existing lessons
Failures traces closed result: fail appended to the project's FAULTS.md with retries + last note
Successful workflows every closed trace: intent, pipeline, retries, transition-timestamp durations 90_META/experience/workflows.json — per-intent runs, pass rate, retries, seconds (the known-good recipes)
Reusable code ast: function bodies across scripts/, difflib > 0.75 (main/selftest excluded — those are convention, not duplication) harvest report: extract-to-shared candidates
Architectural patterns motif regexes counted across the corpus (selftest-guard, docstring-contract, data-table-config, artifact-dir, argparse-subcommands, fail-loud, shared-kernel); ≥3 scripts = confirmed 90_META/experience/patterns.json with evidence

3. Semantic retrieval

recall "<query>" ranks everything harvested — lessons (direct 30_LESSONS/ scan, so drafts surface before any reindex), workflows, patterns — by stemmed-jaccard + difflib similarity, kind-tagged with citations. Empty → "not in vault".

4. Integration

  • /sleep step 6.5: experience.py harvest — learning is part of going to sleep, not optional.
  • /task step 6: after close, harvest that trace immediately (harvest --trace <file>).
  • First real harvest (2026-07-11): 9 traces → 1 lesson drafted (classifier prefix-match fix), all 7 motifs confirmed, and two genuine duplicate functions found (load in query/review/promptc, nwords/norm in experience/skills) — refactor candidates for a future task.