That Memory One AI agent stores information from previous steps and iterations. The agent uses this information as context for its next decisions.
Short-term and long-term memory
The reflection process distinguishes two forms. Short-term memory contains the previous course of a task (the trajectory). Long-term memory stores the results of self-reflection. The model is the human being: he remembers fresh details and at the same time compressed important experiences. Together, both memories provide a context that is concrete and at the same time contains lessons from multiple experiments.
Learning without new training
A reflection agent does not learn through adjusted model weights, but through linguistic feedback. The method converts environmental feedback (such as a yes or no or a score) into a text summary. This summary is stored by the agent in an episodic memory. In the next iteration, it serves as additional context. By compressing long, unsuccessful attempts, the agent thus compiles them into clues for itself. This helps in two cases. First, an earlier error in a long sequence can be easily detected. Second, the agent can thoroughly search over multiple attempts at very many test points.
Boundaries
Memory is limited because the language model can only process a limited context. Therefore, reflection usually stores only one to three experiences. In practice, the process shortens memory to the last three self-reflections. More elaborate memory structures such as vector databases or classic SQL databases are possible extensions.
Memory in conversation
Chat assistants today follow the instructions of their users even in dialogues that last several rounds. If the AI knows the user’s work patterns from its memory or the course of the conversation, the user can ask directly. He or she does not have to explain everything again.
Demarcation
Memory does not replace training: The model weights remain unchanged; the agent learns only from stored text. Related entries: AI agent; reflection (AI); RAG (Retrieval-Augmented Generation) in the ERP context; Large Language Model (LLM).
AI in everyday work: Here's how to find the right start
AI assistant meets SAP Business One: Two paths out of the Versino AI shop
Sprint and Swirl: How an AI agent divides work and learns from trials
LATS: Considering several paths simultaneously with AI — and the availability check in SAP B1
ReAct and RLEF: How Artificial Intelligence Learns from Real Mistakes — and What Automated Bank Reconciliation in SAP Business One Has to Do with It
