ALTK-Evolve: Revolutionizing AI Agent Learning

ALTK-Evolve: Transforming AI Agent Learning

In the field of artificial intelligence, agents are often limited by their inability to learn effectively and apply acquired knowledge to new situations. Traditionally, these agents merely rehash transcripts, leading them to repeat mistakes without being able to transfer lessons learned to other contexts.

ALTK-Evolve stands out by converting the raw experiences of agents into reusable guidelines. During performance tests, this method demonstrated a notable improvement in agent reliability, particularly in complex tasks, with a 14.2% increase on the AppWorld platform, while avoiding overwhelming the working context.

The Problem of the "Perpetual Intern"

Imagine a chef who, despite having memorized all the cookbooks, forgets the specifics of your kitchen every day. He doesn’t remember that your oven runs hotter than normal or that your customers prefer a bit more salt. This chef will follow a recipe to the letter but will be at a loss if there’s a shortage of lemons. This analogy illustrates how many current AI agents operate: they excel at executing precise instructions but struggle to accumulate a genuine understanding of their environment.

A young cook needs distinct recipes for each dish, while an experienced chef understands that "acidity balances fat" and applies this principle universally. Similarly, to be truly effective, AI agents must extract principles from their experiences and apply them to new tasks, rather than merely reproducing past actions. ALTK-Evolve's long-term memory system enables precisely this transformation, converting interactions into candidate guidelines, filtering for quality, and integrating only relevant advice at the right moment.

A recent MIT study revealed that 95% of agent failures are due to their inability to adapt and learn on the job. ALTK-Evolve fills this gap by using long-term episodic memory to enhance agent reasoning.

Solution: Long-Term Memory with ALTK-Evolve

Evolve is a memory system designed to help AI agents continuously improve by learning from and utilizing guidelines derived from previous executions.

The system operates in a continuous loop:

• Downstream (Observation & Extraction): It captures all agent trajectories, including user statements, reflections, tool calls, and outcomes, in an interaction layer. Tools like Langfuse or others based on OpenTelemetry allow for the extraction of structural patterns and their retention as candidate entities.

• Upstream (Refinement & Retrieval): A background consolidation and scoring process merges duplicates, eliminates weak rules, and strengthens proven strategies, thereby creating a library of high-quality guidelines, policies, and standard operating procedures (SOPs). Retrieval selects only relevant items through the interaction layer and reinjects them into the context at the application layer.

This approach is effective for several reasons:

• Teaching Judgment: It transforms isolated events into transferable strategies across different tasks.

• Noise Control: Scoring keeps memory lightweight and useful, avoiding excessive accumulation of irrelevant information.

• Just-in-Time Disclosure: Retrieval occurs just in time, without overwhelming the working context.

Results: Improved Reliability, Especially on Complex Tasks

We tested the framework on AppWorld, where agents perform realistic multi-step tasks via APIs, using an average of 9.5 APIs across 1.8 applications. Complex cases require a more elaborate control flow. A ReAct agent received task instructions along with the 5 best retrieved guidelines from a previous execution and was tested on an unseen partition. We measured the Scenario Goal Completion (SGC), a strict consistency metric that requires success across variants.

The evaluations revealed several key findings:

• Generalization: The agent improves on unseen tasks, demonstrating that it learns principles rather than merely memorizing recipes.

• Complexity Scale: The more complex the task, the more the agent benefits from learned guidelines, with maximum gains on the most difficult tasks. Complex tasks saw a relative increase of 74% in success, with guidelines aiding navigation through complex control flows.

• Consistency: SGC gains surpassed improvements in raw success rates, reducing unpredictable behavior across scenario variants. The guidelines not only help the agent solve tasks but also assist it in solving them reliably across variants.

Getting Started (Choose Your Path)

There are several ways to integrate ALTK-Evolve into your agent.

• No Code with Claude Code, Codex, and IBM Bob (Lite Mode): Install the plugin in Claude Code:
  • claude plugin marketplace add AgentToolkit/altk-evolve
  • claude plugin install evolve-lite@evolve-marketplace

That's it! The plugin extracts entities from trajectories and stores them as files on your file system. It uses Claude Code hooks for automatic retrieval.

• With a ReAct Agent (Low-Code): Add a single altk_evolve.auto import and enable a flag to emit traces to an Arize Phoenix user interface. Then, synchronize the traces to generate improvement guidelines without changing your current stack. This works with popular LLM clients and agent frameworks (e.g., OpenAI, LiteLLM, and Hugging Face agents), allowing you to maintain your current stack while gaining visibility.

• With CUGA (Pro-Code): We have integrated ALTK-Evolve directly into CUGA via MCP to create a tight, low-overhead learning loop. Before each execution, the MCP tool get_guidelines is called to surface task-specific orientations and reduce trial-and-error. After execution, CUGA returns structured execution traces via save_trajectory, allowing Evolve to learn from what actually happened and improve future advice.

Try It and Share Your Findings

Your agent shouldn't wake up like an intern every morning. This approach helps it learn on the job.

If you’re using Claude Code, Codex, and IBM Bob, try it out in just a few minutes and see how it enhances your agent.

Star the repository; it helps others discover the project and directly guides what we build next.