Imported from AgentToolkit/altk-evolve (AGENTS.md). Install upstream with npx skills add AgentToolkit/altk-evolve. Copyright stays with the author.

What is Evolve?

Evolve is a Python library and service which enables AI agents to improve through self-reflection.

Key Concepts

• Trajectory: A recorded agent conversation
• Entity: Anything which is appropriately stored in a vector database, such as a guideline, policy, or some other knowledge.
• Namespace: Isolated storage for entities
• Conflict Resolution: LLM-based merging of duplicate/conflicting entities
• Guidelines: Instructions intended to assist an agent in completing some task

Architecture Flow

1. Agent completes some task, and the resulting trajectory is automatically saved into a logging framework such as Langfuse or Arize Phoenix.
2. The agent can call the sync MCP function, or the user can manually sync, which causes evolve to process the trajectory and save any generated guidelines.
3. Generated guidelines are stored as entities with conflict resolution applied.
4. Future agents can query the Evolve MCP server to fetch guidelines for similar tasks

Project Directory Tree (Some files omitted for brevity)

.
├── demo (Files used by the Claude Code demo)
│   ├── filesystem
│   └── workdir
├── docs (Data used by README files)
├── explorations (Tangential projects for feeling out future work. Should be avoided unless otherwise prompted.)
│   └── claudecode
├── evolve (Primary Source Root)
│   ├── backend (Entity Database Backend implementations, primarily vector databases)
│   ├── cli (A CLI wrapper over the native Python client)
│   ├── config (All configurations which are derived from environment variables or instantiated as an object)
│   ├── db (A sqlite database for when vector databases are a poor fit for the data)
│   ├── frontend (Interfaces to interact with the backend)
│   │   ├── client (A native Python client which thinly wraps the configured backend)
│   │   └── mcp (An MCP server implementing some high-level methods useful for AI agents)
│   ├── llm (All code that prompts an LLM)
│   ├── schema (All well-defined datatypes used throughout the project)
│   ├── sync (Upstream data sources to be processed and stored in the backend)
│   └── utils (Small reusable code snippets)
├── tests (All tests for evolve)
├── .env.example (Environment variable template file)
└── .env (Environment variables used to configure evolve)

First Time Setup

uv sync && source .venv/bin/activate
cp .env.example .env  # Configure any environment variables, defined in `./altk_evolve/config`
pre-commit install

Development Guidelines

• This project is managed by uv, not python or pip, so any python commands need to go through uv. All dependencies are defined in pyproject.toml.

Testing Instructions

• Run pytest verbosely with the -v flag by default so that you have more context when tests fail.
• Use uv run pytest tests/.../<test_name.py> to run tests individually.
• We use the pytest markers e2e for end-to-end tests, and unit for unit tests, and phoenix to test integration with Phoenix.
• When running uv run pytest it will skip the tests marked with phoenix.
• To run specific markers: uv run pytest -m e2e or uv run pytest -m unit
• To run all tests: uv run pytest -m "e2e or unit or phoenix"

Available Interfaces

• MCP Server: get_entities(), get_guidelines(), save_trajectory()
• CLI: Run evolve --help if details are needed about its subcommands. Available subcommands include namespaces, entities, and sync
• Python Client: EvolveClient() for programmatic access

Coding Standards

• Use Ruff for linting and formatting (configured in pyproject.toml)
• Always run uv run ruff format . and git add -u before committing to avoid pre-commit stash conflicts with ruff auto-fixes
• Write commit messages in the Conventional Commits format expected by python-semantic-release:
  • feat(scope): description — new feature, triggers a minor version bump
  • fix(scope): description — bug fix, triggers a patch version bump
  • perf(scope): description — performance improvement, triggers a patch version bump
  • test(scope): description, chore(scope): description, docs(scope): description, etc. — no version bump
  • Breaking changes: append ! after the type/scope (e.g. feat!:) or add BREAKING CHANGE: in the footer
• All new features need tests (unit + e2e where applicable)
• Use uv to run Python commands, including pip.