01 / TRIGGERS
Rules, not search
A memory names the files, symbols, commands or words it applies to. Only matches are delivered, once per session.
PLUGIN FOR CLAUDE CODE · CODEX AND OPENCODE IN PROGRESS
What your code doesn't say, said at the right moment.
DeltaDictum keeps the decisions, traps and unwritten steps of your project as memories in git, and hands each one to the agent right before the tool call it applies to.
/plugin install dd@deltadictum
DD - 12 memories live in this project. Review UI: /dd:review
> add a test for the store's close() method
● Edit tests/store.test.js
DD - project knowledge (tests/store.test.js):
[TRAP] Close stores in t.after
A failing assertion otherwise leaves the handle open and node --test never ends.
✓ Wrote t.after(() => store.close())
What DeltaDictum does
02 / TIMING
At the action
Before the edit, before the command. A session that never touches those files never pays for it.
03 / CAPTURE
Saved with evidence
After a fix or a correction, the agent proposes a memory. Verified evidence activates it; the rest waits for you.
04 / LOCAL
Shared through git
JSON files in your repo. No account, no model calls, no background process, no dependencies.
TABLE 1 · MEASURED
Full benchmark, Claude Opus 5.5, blind judge.
26 tasks, 3 runs per condition. Change against the agent alone.
| Metric | Agent alone | With documentation | With documentation, not referenced | With DD |
|---|---|---|---|---|
| Tasks resolved | Base95% | +5% | +4% | +5% |
| Code quality | Base92% | +5% | +7% | +5% |
| The project's own requirements met | Base69% | +45% | +42% | +45% |
| Context | Base37.8k tokens | +9% | +7% | +5% |
| Cost per task | Base$0.428 | +17% | +11% | +6% |
Four projects, 26 tasks selected where Haiku alone failed or controls, three runs each; on track, not a promise for yours. All trials and how they are measured →