wtf is open-reasoning-format?

The Open Reasoning Format (ORF) is a specification and reference toolkit that gives AI agents a file-based memory system. Instead of relying on vector databases or external servers, agents store and retrieve procedural knowledge, heuristics, and lessons learned from past tasks as plain Markdown and YAML files in a local workspace folder. The project is currently in draft status at version 0.1.0. The core idea is progressive disclosure. When an agent starts a complex task, it first queries a lightweight category index that costs roughly 200 tokens of context. If a category matches the task domain, the agent inspects metadata for relevant experience records, spending another 500 tokens. Only when a specific trigger condition matches does the agent load a full playbook, around 800 tokens, into its working context. This staged approach keeps context usage low while still making detailed operational knowledge available when needed. Each experience record follows a standardized five-section Markdown structure with YAML frontmatter for indexing and trigger conditions. The repository ships with a Python helper script called experiences.py that provides deterministic file operations: listing categories, inspecting metadata, reading a specific record by ID, and creating new records. Creating a new experience involves specifying a domain, title, description, keywords, complexity level, an objective, a trap the agent fell into, an insight gained, a validated solution path, and a checklist item for verification. ORF also includes a pre-built Agent Skill called manage-experience that conforms to the agentskills.io specification. Compatible agent frameworks can discover this skill automatically. The skill guides agents through a two-phase lifecycle: retrieval of past experiences before starting a task, and recording of new learnings after successfully resolving a problem. This means an agent that figures out a tricky workaround can persist that knowledge for future runs. The repository bundles an evaluation harness under the evals directory. This framework runs trajectory benchmarks to measure whether experience playbooks actually improve agent performance by reducing step counts and eliminating trial-and-error debugging cycles. The evals include a sandbox runner, spec validator, isolated scenarios, and exported reports. The full README is longer than what was shown.