irzix/experia

Experia is an open-source Python library that adds a learning layer to AI agents. Most AI agents start fresh every time they run a task, which means they repeat the same mistakes. Experia tries to fix this by capturing what an agent did, evaluating whether it worked or failed, extracting a lesson, and feeding that lesson back into future decisions. It works alongside existing agent frameworks like LangChain and LangGraph rather than replacing them. The project describes its learning loop as a cycle: observation, action, result, experience, lesson, memory, and then a better future action. Internally, Experia stores actions in a shared memory, uses a language model to evaluate outcomes, generates rules from those evaluations, and can run a reflection step that looks at batches of past actions to produce broader strategies. When an agent is about to act, Experia retrieves relevant context from this memory and injects it so the agent benefits from prior experience. The README includes a benchmark designed to show whether this actually helps. Two identical agents face the same set of operational tasks across four rounds. The baseline agent succeeds zero times out of 24 attempts and repeats the same avoidable mistake 18 times. The Experia agent fails each task at most once, learns from the failure, and reaches a 100% success rate from the second round onward. This benchmark runs fully offline with no API key required. Installation is through pip, with optional extras for language model features and specific integrations such as LangGraph or OpenAI. The README provides code examples for both a native LangGraph integration and a manual core API. In the manual mode, you create a learner, record actions and results, and optionally trigger a reflection step. Retrieval can be semantic if you provide an embedder, otherwise it falls back to keyword search. The full README is longer than what was shown.