graphify 0.8.47
Self-improving "work memory" (#1441)
graphify save-result --outcome useful|dead_end|corrected [--correction TEXT]records how a saved Q&A turned out.graphify reflectaggregatesgraphify-out/memory/into a deterministicreflections/LESSONS.mdan agent loads next session. Source nodes are scored, not counted — signed and recency-decayed (configurable--half-life-days, default 30), so a fresh dead end outweighs a stale useful. A node is preferred only once corroborated by ≥--min-corroborationdistinct results (default 2); others are tentative; mixed-signal nodes render once as contested (recency-wins). Citations whose node no longer exists in the graph are dropped, so stale lessons don't linger. Deterministic, no LLM.- Zero-config adoption: the skill instructs the agent to read
LESSONS.mdat the start of graph work and to record outcomes, and the git post-commit/post-checkout hooks now auto-runreflectafter each rebuild — so lessons stay fresh without a manual command.
Fixes
- Python qualified class-method calls (
ClassName.method(...)) now produce an EXTRACTEDcallsedge to the class-qualified method node (#1446) — previously dropped, including the common case where the called method shares its name with the caller (a viewset action delegating to a same-named service action). validate_extraction/build_from_jsonno longer crash on a non-hashable nodeidor edge endpoint from a malformed extraction — a single bad node no longer aborts the whole build (#1447, thanks @dschwartzi).graphify updatenow prunes a function/symbol removed from a still-present file without--force— the shrink-guard is file-aware (allows shrinkage from re-extracted/deleted files, still refuses unexplained loss).
Install: uv tool install graphifyy==0.8.47