IBM Research · 2026-04-08 · notable
ALTK-Evolve — continuous learning for AI agents
IBM Research open-sources a long-term memory system that turns agent trajectories into reusable guidelines — so an agent can learn from its own runs instead of re-reading transcripts.
An open-source long-term memory layer for agents: turn what your agent did yesterday into guidelines it actually follows tomorrow.
What is it?
ALTK-Evolve is an open-source component of IBM Research's Agent Toolkit that gives AI agents a form of on-the-job learning. Instead of stuffing a model's context window with past transcripts, ALTK-Evolve converts each agent trajectory into short, reusable guidelines and injects only the relevant ones at the moment of action.
How does it work?
The system runs as a loop. As an agent operates, ALTK-Evolve extracts candidate guidelines from its trajectories, consolidates and scores them for quality, stores them in a vector index backed by Milvus, and uses an LLM-based conflict-resolution step when new guidelines contradict existing ones. An MCP server then surfaces the most relevant guidelines to the agent at runtime.
Why does it matter?
Agents still forget whatever is not in their immediate context. ALTK-Evolve is a clean, Apache-2.0-licensed implementation of a structured middle ground between 'dump the whole log' and 'fine-tune the model.' For teams running agents in production, this is the kind of infrastructure that makes 'it gets better the more you use it' real.
Who is it for?
Agent framework authors, production agent teams.
Try it
github.com/AgentToolkit/altk-evolve