Last month my team reversed a queue decision. Nobody could say why we
had picked the first design, what evidence it rested on, or who had
signed off on the replacement. The old design was still half-wired in
as a "fallback", quietly depending on the thing we had just replaced.
We found out in production.
Vector RAG remembers text. It does not remember decisions — what
was chosen, what it rested on, what contradicted it, and what replaced
it. So I built a tiny local graph for exactly that: memgraph.
One zero-dependency Node.js CLI. No network, no server, no embeddings.
The store is a single JSON file you can commit next to your code.
memgraph init
memgraph add --type decision --title "Use Postgres LISTEN/NOTIFY for the queue" --source "ADR-014"
memgraph add --type fact --title "NOTIFY payloads cap at 8000 bytes" --source "Postgres docs"
memgraph link m-002 m-001 --relation contradicts
memgraph add --type decision --title "Switch the queue to Redis Streams" --source "ADR-019"
memgraph supersede m-001 m-003 --reason "payload cap makes NOTIFY the wrong tool"
Three questions it answers, fast:
memgraph why m-003 — source,
supporters, contradictions, supersession.memgraph check —
unresolved contradictions, live nodes resting on superseded ones,
sourceless claims.memgraph list — every node, with
superseded ones marked.
memgraph check exits 1 when it finds an error, so it works as a CI
gate: a PR that introduces a contradiction, or keeps depending on a
superseded decision, fails before it merges.
The repo ships a small team graph — a queue decision, the fact that
killed it, its replacement, and one leftover fallback still depending
on the dead design:
sh examples/run-demo.sh
memgraph check on that graph reports exactly two problems: the
fallback reader depends on a superseded decision, and one open
question has no source. The contradiction between the payload-cap
fact and the old decision does not get flagged — the old decision
was superseded, so that contradiction is resolved history, not a
live conflict.
A graph is only as good as what you record. memgraph cannot tell you
whether a source is true — it tells you where every claim came from
and whether the graph still holds together. A decision with no source
is a rumor with formatting; that is why check flags it. And when a
decision changes, supersede it — don't delete it. Deleting history is
how teams end up re-litigating the same choice every six months.
Repo: https://github.com/techdevfun/memgraph — MIT, zero dependencies,
Node.js 18+.