personal-evotaste 0.1.0

Developer personal-taste memory & self-evolution engine for AI coding agents.

🧬 PersonalEvoTaste

Teach your AI coding agent your taste — and watch it evolve.

---

✨ What is it?

LLM agents are great at writing generic code — but they have zero memory of your personal preferences: indentation tastes, library choices, naming conventions, the brutalist UI you love, the verbose docstrings you hate. Every new chat resets to zero.

PersonalEvoTaste is a small, dependency-light library that gives an AI agent a persistent, self-evolving memory of your taste. Feed it the agent's output + your feedback, and it distils, deduplicates, reinforces and decays rules over time — then re-injects the top ones into the next prompt.

        ┌──────────────┐  feedback   ┌──────────────────┐  rules   ┌────────────┐
agent → │ your review  │ ──────────► │ PersonalEvoTaste │ ───────► │ next prompt│
        └──────────────┘             └──────────────────┘          └────────────┘
                                              ▲  decay/reinforce        │
                                              └─────────────────────────┘

🚀 Features

• 🧠 Typed memory — Pydantic v2 models, JSON-schema-friendly, versioned on disk.
• 🔌 Pluggable storage — YAML / JSON / SQLite out of the box; bring-your-own backend.
• 🔬 Pluggable extractors — deterministic heuristic by default; plug an LLM with CallableExtractor.
• ♻️ Self-evolution — automatic deduplication, reinforcement on repetition, time decay.
• ⚖️ Ranked injection — top-N rules by weight, optionally filtered per project.
• 🖥️ Zero-dep CLI — personal-evotaste evolve / inject / list / remove / reset / export.
• 🧪 Tested & typed — pytest suite, ruff, mypy, CI on Python 3.9 → 3.12.
• 💾 Crash-safe writes — atomic file replacement, SQLite transactions.

📦 Installation

pip install personal-evotaste
# or from source
pip install -e ".[dev]"

⚡ Quickstart

from personal_evotaste import PersonalEvoTaste

taste = PersonalEvoTaste(memory_path="my_taste.yaml")

# 1. Inject your evolving taste into any prompt
prompt = taste.inject_taste(
    "Generate a login page",
    context="brutalist minimal project",
)
# → send `prompt` to your LLM of choice

# 2. After reviewing the agent's output, teach it
taste.evolve(
    agent_output="A flashy gradient login page with neon buttons",
    user_feedback="Too loud. I want calmer, monochrome, brutalist typography.",
    project_name="acme-landing",
)

# 3. Next time, the rule is already in the prompt — and reinforced if you repeat yourself
print(taste.inject_taste("Generate a signup page"))

🖥️ CLI

# Add a rule from feedback
personal-evotaste evolve \
  --output "A flashy gradient login page" \
  --feedback "Too loud. Prefer calm monochrome brutalist typography." \
  --project acme-landing

# Inject taste into a prompt
personal-evotaste inject "Generate a signup page" --context "acme-landing"

# Inspect the memory
personal-evotaste list
personal-evotaste export > snapshot.json

Memory backend is auto-selected from the file extension: .yaml / .json / .sqlite3.

🔌 Plugging an LLM extractor

from personal_evotaste import PersonalEvoTaste, CallableExtractor

def summarise_with_llm(agent_output: str, feedback: str, project: str) -> str:
    # call OpenAI / Anthropic / a local model here and return one short rule
    ...

taste = PersonalEvoTaste(
    memory_path="my_taste.yaml",
    extractor=CallableExtractor(summarise_with_llm),
)

🧱 Architecture

personal_evotaste/
├── models.py        # Pydantic models: TasteRule, FeedbackEvent, TasteMemory
├── storage.py       # BaseStorage + YAML / JSON / SQLite backends (atomic writes)
├── extractors.py    # RuleExtractor strategies + similarity-based deduplication
├── core.py          # PersonalEvoTaste façade (evolve / inject / ranking / decay)
├── cli.py           # argparse-based CLI (zero extra deps)
├── exceptions.py    # Typed error hierarchy
└── logging_config.py

See docs/architecture.md for the design rationale.

🤝 Contributing

PRs, issues and ideas are very welcome — see CONTRIBUTING.md and our Code of Conduct. If PersonalEvoTaste helps you, a ⭐ on GitHub goes a long way.

📜 License

MIT © PersonalEvoTaste Contributors. See LICENSE.