New in 0.8.47: self-improving "work memory" so graphify learns from how you use it

[safishamsi]

graphify now keeps a small "work memory" of how its answers turned out, and brings those lessons back at the start of your next session. The idea is simple: the more you use it, the better it gets at pointing you to the right places. This landed in 0.8.47. It is deterministic, uses no LLM, and stays quiet until you actually record an outcome.

The basic loop

1. When you save an answer, say how it turned out:

graphify save-result --question "how does auth work?" --answer "JWT via verify_token()" \
  --nodes "verify_token()" "jwt_auth.py" --outcome useful

--outcome takes one of three values:

• useful: the nodes you cited actually answered the question, so they get treated as preferred sources later
• dead_end: the question or path went nowhere, so there's no point re-deriving it
• corrected: the answer was wrong. Pass --correction "the right answer" to record what it should have been

2. Run reflect to fold those outcomes into a lessons file:

graphify reflect            # writes graphify-out/reflections/LESSONS.md

3. Next time, read LESSONS.md before you start. It tells you which sources tend to be worth starting from, which questions have been dead ends, and what got corrected.

You usually won't run any of this by hand

If you work through a coding agent, the loop is already set up for you:

• The skill tells the agent to read LESSONS.md at the start of graph work, and to record an outcome when it saves an answer.
• The git post-commit and post-checkout hooks run reflect for you after each rebuild (only when you have outcomes saved), so the lessons file stays current without you thinking about it.

Run graphify hook install once and that's it.

How the lessons get ranked

It scores citations rather than just counting them:

• Recency matters. Each citation fades over time (tune it with --half-life-days, default 30), so a recent dead end carries more weight than a "useful" from months ago.
• A node has to earn it. It only becomes a preferred source once it shows up in at least two separate useful results (--min-corroboration, default 2). A single save shows up as tentative, not trusted.
• Conflicts are shown honestly. If a node was useful once and a dead end later, it appears once with a "most recent wins" note instead of quietly sitting in both lists.
• It keeps up with your code. If a citation points at a node that no longer exists in the graph, it gets dropped, so deleting code clears its old lessons (a normal graphify update is enough).
• Same inputs give you the same LESSONS.md every time.

Try it

uv tool install graphifyy==0.8.47

This is a first, deliberately simple version (#1441). A few things are still on the list: richer LLM-written lessons, separating "judge how it went" from "save the answer", and pulling only the lessons relevant to the current question. If you give it a run on your own codebase, I'd love to hear how it holds up.