Memento-Skills Repository Summary

Memento-Skills is a Python 3.12+ agent runtime and research codebase that implements the paper’s self-evolving skill-memory design. The repository is not only a benchmark release: it includes the runtime, configuration system, local sandboxing, verification commands, built-in skills, and multiple operator surfaces.

Core repository shape

• core/skill/ separates skill building, downloading, loading, retrieval, storage, and execution into explicit subsystems.
• core/memento_s/phases/ and core/protocol/ expose a staged agent pipeline rather than a single monolithic loop.
• middleware/sandbox/ plus execution-policy modules (tool_gate, path_validator, pre_execute, recovery) make tool use an explicit control layer.
• builtin/skills/*/SKILL.md packages skills as markdown artifacts with metadata, scripts, and optional supporting assets.
• tests/ covers skill retrieval, execution, downloading, storage, schema handling, and tool-security paths.

Surfaces and operations

• CLI entrypoint: memento
• Desktop GUI entrypoint: memento-gui
• Verification command: memento verify
• Messaging and gateway integrations under im/ and middleware/im/gateway/

Architectural signal

The repository makes the paper’s thesis concrete: skills are not static prompts pasted into context but durable packages that can be retrieved, executed, verified, downloaded, and rewritten over time. The codebase therefore belongs in this wiki as a harness artifact, not only as a benchmark attachment.