Experia AI

Memory layers help your agents remember. Experia helps them learn.

The open-source experience-learning layer that makes AI agents measurably stop repeating their mistakes. Experia captures what an agent did, evaluates why it worked or failed, distills a reusable lesson, and reinforces the strategies that keep paying off — a cognitive plugin for the frameworks you already use (LangChain, LangGraph, and more). It doesn't replace your agent framework; it closes the feedback loop around it.

Vision

Current AI agents start from zero on every interaction. Experia adds an experience learning loop around agents. It allows agents to remember what happened, understand why it worked or failed, and improve future decisions.

Observation → Action → Result → Experience → Lesson → Memory → Better Future Action

flowchart TD
    subgraph Multi-Agent Swarm
        AgentA[Coder Agent]
        AgentB[Researcher Agent]
        Supervisor[Supervisor Agent]
    end

    subgraph Experia AI Cognitive Layer
        Store[(Shared MemoryStore)]
        Eval[LLM Evaluator\nRoot Cause Analysis]
        RuleGen[Rule Generator]
        Reflect[Reflection Engine\nBatch Analysis]
        Ctx[Context Builder]
        
        AgentA & AgentB -- 1. Record Action --> Store
        Store -- 2. Evaluate Outcome --> Eval
        Eval -- 3. Extract Lesson (agent_role) --> Store
        Eval -- 4. Consolidate --> RuleGen
        RuleGen -- 5. Generate RULE --> Store
        Supervisor -- 6. Trigger reflect() --> Reflect
        Reflect -- 7. Generate Global STRATEGY --> Store
        Store -- 8. retrieve_context(agent_role) --> Ctx
    end
    
    Ctx -- 9. Inject Shared Knowledge --> AgentA & AgentB & Supervisor

Does it actually work?

A deterministic, fully offline benchmark (benchmark/learning_benchmark.py) pits two identical agents against the same workday of ops tasks. The agent has no built-in knowledge — its only intelligence is the context Experia injects, so every difference is attributable to the cognitive layer.

Experia learning benchmark  (6 tasks x 4 rounds = 24 episodes)
==================================================================
Metric                                  Baseline       Experia
------------------------------------------------------------------
Tasks completed successfully               0/24         18/24
Overall success rate                          0%           75%
Mistakes repeated (avoidable)                 18             0
==================================================================

Success rate per round (the learning curve)
------------------------------------------------------------------
  Round 1   baseline [....................]   0%   experia [....................]   0%
  Round 2   baseline [....................]   0%   experia [####################] 100%
  Round 3   baseline [....................]   0%   experia [####################] 100%
  Round 4   baseline [....................]   0%   experia [####################] 100%
------------------------------------------------------------------

The baseline agent makes the same avoidable mistake 18 times. The Experia agent fails each task at most once, extracts the lesson, and never repeats it — reaching a 100% success rate from the second encounter on. No LLM or API key is required to reproduce this:

python benchmark/learning_benchmark.py

Integrations

Experia acts as a cognitive plugin. It does not replace your agent frameworks (like LangChain, AutoGen, CrewAI), it enhances them by managing long-term memory, learned experiences, user knowledge, and behavioral patterns.

Getting Started

Install the base package for the offline core quickstart:

pip install experia

Optional features have separate extras: experia[llm] for LiteLLM-backed evaluation/embedding/rules/reflection, experia[langchain] for LangChain, and experia[langgraph] for LangGraph. Provider credentials are needed only when the selected LLM or embedding provider requires them.

For class and method details, see the API Reference.

Offline quickstart

This is the canonical executable quickstart from examples/quickstart.py. It needs no network access, credentials, or optional extra, and its assertions document the installed ExperienceRecord and Memory behavior.

"""Executable, offline Experia quickstart using only the base installation."""

import asyncio

from experia import Learner, MemoryType, SimpleHeuristicEvaluator, SQLiteStore


async def main() -> None:
    store = SQLiteStore(":memory:")
    await store.initialize()
    try:
        learner = Learner(
            store=store,
            evaluator=SimpleHeuristicEvaluator(),
        )

        experience = await learner.record(
            task="Deploy web app",
            action="Restart Nginx",
            result="failed with config syntax error",
            context={"attempt": 1},
        )
        assert experience.task == "Deploy web app"
        assert experience.action == "Restart Nginx"
        assert experience.result == "failed with config syntax error"
        assert experience.context == {"attempt": 1}

        persisted = await store.get_experience(experience.id)
        assert persisted is not None
        assert persisted.model_dump() == experience.model_dump()

        await learner.flush()
        memories = await store.search_memories(memory_type=MemoryType.LESSON)
        assert len(memories) == 1
        lesson_memory = memories[0]
        assert lesson_memory.type is MemoryType.LESSON
        assert lesson_memory.agent_role == "default"
        assert lesson_memory.confidence == 0.6
        assert lesson_memory.source == f"experience_{experience.id}"
        assert "Restart Nginx" in lesson_memory.content

        reinforced = await learner.reinforce(lesson_memory.id, success=True)
        assert reinforced is not None
        assert reinforced.reinforcement_count == 1
        assert reinforced.success_count == 1
        assert reinforced.confidence == 0.68
    finally:
        await store.close()


if __name__ == "__main__":
    asyncio.run(main())

Run the same checked source with:

python examples/quickstart.py

Learner always requires both store and evaluator; the quickstart uses the base-package SimpleHeuristicEvaluator. Network-backed calls are kept out of the offline quickstart. The installed constructor examples below exercise each optional import without making network requests; the API gate runs each script in an environment containing only its declared extra.

• LLM evaluation, embeddings, and rules: experia.experience.llm_evaluator.LLMEvaluator, experia.LiteLLMEmbedder, and experia.improvement.rules.RuleGenerator with experia[llm] (example).
• LangChain callbacks/retrieval: experia.integrations.langchain.callbacks.ExperiaCallbackHandler and experia.integrations.langchain.retrievers.ExperiaLearningRetriever with experia[langchain] (example).
• LangGraph nodes: experia.integrations.langgraph.nodes.ExperiaContextNode and experia.integrations.langgraph.nodes.ExperiaLearningNode with experia[langgraph] (example).

The exact example-to-extra mapping is machine-readable in examples/installed-examples.json.

Project Status

Implemented today

Each implemented entry links to an executable example or automated executable test that exercises the installed behavior.

Planned (unavailable in the current version)

The following roadmap items are not implemented in version 0.8.0 and are not used by any quickstart. Existing placeholder imports fail explicitly with UnavailableFeatureError rather than presenting an operational backend. Each item's owner/team/unassigned status and readiness are tracked in the machine-validated roadmap-ownership.yml manifest:

• AutoGen native integration (integration) — readiness planned, ownership unassigned; no importable placeholder yet.
• CrewAI native integration (integration) — readiness planned, ownership unassigned; placeholder experia.integrations.crewai.CrewAIIntegration raises UnavailableFeatureError.
• Distributed production mode (Redis-backed queue + workers) (runtime) — readiness planned, ownership unassigned; no importable placeholder yet.
• Mem0 memory adapter (adapter) — readiness planned, ownership unassigned; placeholder experia.adapters.mem0.Mem0Adapter raises UnavailableFeatureError.
• OpenAI Agents native integration (integration) — readiness planned, ownership unassigned; no importable placeholder yet.
• PostgreSQL + pgvector store (adapter) — readiness planned, ownership unassigned; placeholder experia.adapters.postgres.PostgresAdapter raises UnavailableFeatureError.
• Zep memory adapter (adapter) — readiness planned, ownership unassigned; placeholder experia.adapters.zep.ZepAdapter raises UnavailableFeatureError.

Contributions toward these are welcome — see CONTRIBUTING.md.

License

This project is licensed under the MIT License. See the LICENSE file for details.