ALTK-Evolve: On-the-Job Learning for AI Agents

ALTK-Evolve is a long-term memory system that, instead of having an AI agent re-read the conversation history every time, extracts principles from execution traces and makes them immediately usable for the next task. Rather than simply re-injecting the transcript, it organizes experience into guidelines for reuse.

The core flow has two stages.

• Downward flow: From the full trajectory accumulated in the Interaction Layer (an OpenTelemetry-based observability tool like Langfuse), an extractor pulls out structural patterns and stores them as candidate entities.
• Upward flow: A background consolidate-and-score job merges duplicates, removes weak rules, and reinforces validated strategies to build a high-quality rule library.

This refined knowledge is re-injected only when needed. As a result, the agent can gradually accumulate context-appropriate judgment criteria without excessively bloating its context.

In evaluation, the effect was confirmed on AppWorld. When the ReAct agent was given the top 5 guidelines generated from prior runs and Scenario Goal Completion (SGC) was measured, Easy improved from 79.0% → 84.2%, Medium from 56.2% → 62.5%, Hard from 19.1% → 33.3%, and the overall score from 50.0% → 58.9%.

The effect was especially large on harder tasks. The Hard segment rose by +14.2%p, a relative success rate increase of 74%. The authors explain that this result shows the system is learning transferable principles rather than simple memorization, and that it also increases consistency across scenario variants.

There are also three adoption paths.

• No-code: Apply directly via plugins for Claude Code, Codex, and IBM Bob
• Low-code: Add altk_evolve.auto to export traces and generate guidelines
• Pro-code: Integrate into CUGA via MCP, closing the learning loop with get_guidelines before execution and save_trajectory after execution

Ultimately, the message is clear: agents need to move away from a structure where they start from scratch every morning like an intern, and toward systems that get better with every execution.