altk-evolve
View on GitHubSelf improving agents through iterations
Evolve is an agent self-improvement system that stores conversation trajectories as episodic memory, extracts guidelines, and feeds them back so agents stop repeating mistakes. Ships an MCP server, vector backends, and plug-ins for Claude Code and Codex.
Use Cases
Give coding agents persistent cross-session memoryExtract reusable guidelines from agent trajectoriesRetrieve task-relevant guidelines before each runResolve conflicting learned guidelines with an LLMShare and publish learned entities across agent namespacesRedact PII from agent memory via hooksEnforce retention policies on agent knowledge storesAttach Evolve memory to Claude Code, Codex, or IBM Bob via MCP
Built With
- Language
- Python
- Frameworks
- LiteLLM · FastMCP · FastAPI · Arize Phoenix · sentence-transformers · Milvus · pgvector · OpenInference · smolagents · OpenAI Agents SDK · Typer · Uvicorn · Jinja2 · Pydantic
Tags
agent-memory · self-improving-agents · episodic-memory · trajectory-learning · mcp-server · claude-code · codex · context-management · vector-store · conflict-resolution · guidelines · observability · agent-learning · pii-redaction · retention-policy · llm