Quoth v3.3.0 — Self-Learning as a Service

Quoth v3.3.0 is the biggest release since v3.0. The headline: any developer can now install the self-learning plugin without needing their own AI API keys. Behind the scenes, we rebuilt the daemon's query path, switched embedding providers, and opened the door to cross-organization pattern sharing.

Managed Mode — Zero-Config Self-Learning

Until now, getting Quoth's self-learning loop running meant configuring at least two API keys: AI_GATEWAY_API_KEY for embeddings and LLM calls, plus MOONSHOT_API_KEY if you wanted the Kimi distiller. For teams that just want to learn from their trajectories, that was too much friction.

Managed mode removes all of it.

node cli.js init

The CLI detects your environment and sets the mode automatically:

$ node cli.js init

  Quoth v3.3.0 — Self-Learning Setup

  Detected: Claude Code CLI, no AI keys configured

  ? Select mode:
    ❯ Managed (recommended) — we handle the AI pipeline
      Self-hosted — bring your own keys

  ✓ Mode set to managed
  ✓ Daemon configured → trajectories sync to Quoth cloud
  ✓ Patterns will be delivered within ~60 seconds of session end

  Done. Start a new Claude Code session to begin learning.

With QUOTH_MODE=managed, the daemon sends raw trajectories to our cloud infrastructure. We run the full JUDGE → DISTILL → CONSOLIDATE pipeline on our side and push distilled patterns back to your local memory.db. Your code never leaves your machine — only tool-call metadata and file paths are transmitted.

The cost works out to roughly $0.03/month per active user, covered by the Quoth subscription. No per-token billing, no surprise invoices.

10x Faster Pattern Injection

Every time you type a prompt, Quoth's UserPromptSubmit hook needs to look up relevant patterns. In v3.2, that meant loading the SQLite database, initializing the ONNX runtime for MiniLM embeddings, and querying the HNSW index — all inside a short-lived hook process. Total overhead: ~200ms.

In v3.3.0, the daemon runs a persistent Unix socket server. Hooks send a single HTTP request over the socket and get patterns back in ~20ms.

curl --unix-socket ~/.quoth/daemon.sock http://localhost/health
# {"status":"ok","pid":240990,"uptime":9.15}

The architecture is simple: one long-running process holds the database connection, the loaded ONNX model, and the HNSW index in memory. Hooks connect, query, disconnect. No cold starts, no repeated initialization.

Pattern queries follow the same socket path:

curl --unix-socket ~/.quoth/daemon.sock \
  -X POST http://localhost/query \
  -H "Content-Type: application/json" \
  -d '{"text":"vitest mock patterns","limit":3}'
# [{"name":"mock-first-testing","confidence":0.82,"action":"..."}]

The daemon starts automatically on SessionStart and shuts down gracefully on SessionEnd. If it crashes, the hooks fall back to direct DB access — same as v3.2, no degradation.

Venice Embeddings & Cost Optimization

Cloud-side embeddings have moved from Voyage AI's voyage-4-lite ($0.02/MTok) to Venice BGE-M3 — a multilingual model with comparable quality at a fraction of the cost. This matters for managed mode, where we run embeddings on every trajectory batch.

Local embeddings remain unchanged: MiniLM-L6 via ONNX runs entirely on your CPU at zero cost with ~5ms latency per embedding. No API calls, no network dependency.

For the daemon pipeline, we added a batch embedding endpoint that amortizes overhead across trajectory chunks:

const embeddings = await batchEmbed(chunks, {
  model: "bge-m3",
  batchSize: 64,
});

The total AI cost per active user in managed mode comes out to approximately $0.003/day — embeddings, Haiku distillation, and pattern consolidation included.

Cross-Org Pattern Sharing

Some patterns are universal. "Always validate environment variables at startup." "Use structured error types instead of string messages." "Run tests before committing." These should not need to be re-learned by every team.

v3.3.0 introduces shared patterns — an opt-in system where high-confidence patterns can be contributed to a public pool:

POST /api/v1/patterns/shared
Content-Type: application/json

{
  "pattern_id": "uuid",
  "org_id": "anonymous",
  "min_confidence": 0.8
}

Patterns eligible for sharing must have:

• Confidence score above 0.8
• At least 10 successful applications
• No project-specific file paths or variable names
• Category marked as broad (not domain-specific)

When your daemon pulls shared patterns, they start at 0.6 confidence locally. They have to prove themselves in your codebase before reaching injection threshold. Bad patterns decay naturally through Quoth's Bayesian scoring.

The long-term vision is a pattern marketplace — curated collections for specific stacks, frameworks, and workflows. v3.3.0 lays the infrastructure.

Habla Espanol

Quoth's task router now understands 20+ categories in Spanish, with accent-agnostic matching. Tildes are optional — the router normalizes input before classification.

"arregla este bug"           → coder@0.8
"disena la landing page"     → designer@0.7
"revisá los tests"           → tester@0.8
"desplegá a producción"      → deployer@0.9

This covers Argentine voseo, Mexican formal, and everything in between. The keyword patterns live in routing.js alongside the English equivalents — no separate language module, no translation layer.

What's Next

We are working toward making Quoth installable as an npm package:

npx quoth init

One command, zero configuration, self-learning active. Beyond that, the roadmap includes a usage dashboard showing pattern hit rates and confidence trends, and the full pattern marketplace for sharing learned intelligence across teams.

---

Quoth v3.3.0 is available now. Run node cli.js init in the quoth-plugin directory to get started, or update your existing installation with git pull && bash scripts/setup.sh.