creation.
← ideas
llm llmgate

Behavior-Learning Memory as an Agent-Infra Primitive

Store how an agent's behavior changed in response to user corrections as a first-class, versioned, replayable artifact — not facts.

memoryagentsdspylanggraphpreference-learning
ANGLES §14. Mem0/Graphiti/Zep/Letta all store facts; none store behavior-deltas. Customers: DSPy/LangGraph/Pydantic AI operators running fast-iteration agents, vertical AI startups whose differentiation is "our agent learns your preferences," agent-infra vendors (LangChain, Arize, Langfuse). Deep-tech: a behavior-diff representation that is trace-reproducible, replay-testable under model upgrades, composable across agents. Kill: Anthropic/OpenAI ships an API-level preference-snapshot primitive in Agents SDK/MCP.

source: ANGLES.md#14

Signals connected to this idea

No signals have connected here yet.