AI that learns your judgment. Gradata captures your corrections to AI output, extracts behavioral instructions, and graduates them into rules that auto-inject into future sessions. Over time, the AI converges on how you think -- not generally smarter, but calibrated to you.
# Clone into your Claude Code plugins directory
git clone https://github.com/gradata-ai/gradata-plugin ~/.claude/plugins/gradata
# Install the Python SDK (required for the daemon)
pip install gradataVerify the plugin loaded:
/gradata doctor
Gradata runs a three-stage graduation pipeline. Corrections start weak and strengthen through repetition:
Session 1: You correct Claude's output (change em dash to comma in an email)
-> Gradata extracts: "Never use em dashes in email prose"
-> Lesson created as INSTINCT (confidence 0.40)
Session 3: Same correction again
-> Confidence boosted to PATTERN (0.60)
Session 7: No more em dash corrections needed
-> Graduated to RULE (confidence 0.90)
-> Auto-injected into every future session
Rules that stop being useful decay. Rules that conflict get flagged. The system self-corrects.
+-------------------------------------------+
| Claude Code |
| |
| SessionStart --> inject rules |
| UserPrompt --> scope-match + detect |
| Edit --> capture correction |
| Stop --> graduation sweep |
| | |
| v |
| localhost HTTP daemon (Python) |
| | |
| v |
| ~/.gradata/projects/<hash>/ |
| lessons.md | system.db | events.jsonl |
+-------------------------------------------+
The plugin communicates with a local Python daemon over HTTP. All processing happens on your machine. The daemon manages the brain vault (lessons, rules, events) per project.
| Command | Description |
|---|---|
/gradata status |
Show brain health: rule count, lesson stats, session number |
/gradata doctor |
Diagnose daemon, config, and plugin connectivity |
/gradata review |
Review pending lessons and promote/reject them |
/gradata promote |
Manually promote a lesson to a higher confidence tier |
/gradata forget |
Remove a lesson or rule by ID |
/gradata prove |
Generate a provenance proof for the current brain state |
The plugin detects corrections through multiple channels:
- Explicit corrections -- edits to AI-generated output (diffs tracked by severity)
- Implicit feedback -- phrases like "that's wrong", "stop doing X", "I told you before"
- Acceptance signals -- a rule fires and the output is not corrected (reinforcement)
- Addition patterns -- repeatedly adding the same thing (type annotations, imports, headers)
- Context switching -- different behavior expected in code vs email vs config
- Correction conflicts -- a new edit contradicts a recent lesson (flags for review)
"Daemon not available"
The Python daemon is not running or unreachable. Run /gradata doctor to diagnose. The daemon should auto-start on session begin.
"No rules injecting"
Rules require graduation. A correction must repeat across 3+ sessions to reach PATTERN, and further to reach RULE. Check /gradata status for pending lessons.
"Wrong Python"
The daemon needs the Python environment where gradata is installed. Check ~/.gradata/config.toml and update python_path to point to the correct interpreter.
"Plugin not loading"
Verify the plugin directory contains .claude-plugin/plugin.json. Run ls ~/.claude/plugins/gradata/.claude-plugin/ to confirm.
- All data stays local in
~/.gradata/ - The daemon binds to
127.0.0.1only -- no network exposure - No cloud sync in v1
- Optional anonymous telemetry is opt-in and content-free (event counts only)
- Python 3.10+
- Claude Code CLI
gradataPython package (pip install gradata)
AGPL-3.0. See LICENSE.