LangChain: Memory and UX are key to building effective proactive agents
·2026.11.18 01:56
Key point
Over 70% of internal agent runs now come from schedules rather than human prompts, driven by memory systems and Slack-integrated review workflows.
Details
- Proactive agents, which run in the background toward goals rather than waiting for prompts, rely on two essential components to manage human attention effectively: memory and user experience (UX).
- Memory allows agents to learn from seed data, user feedback, and outcomes, enabling them to become more autonomous and reduce the frequency of check-ins over time.
- UX design focuses on high-signal interactions, built-for-review interfaces, and integration into existing workspaces like Slack to minimize friction and maximize feedback loops.
- LangChain's internal agents, GTM Agent and Robin, demonstrate this approach, with over 70% of their runs now triggered by schedules rather than ad hoc human prompts.
Memory Architecture
- Memory sources include seed data (past context), feedback (present user corrections), and outcomes (future results of agent actions).
- Maintaining memory requires handling salience, expiry, and duplication, often via a background job called "dreaming" that updates memory files.
- The effectiveness of memory is measured by metrics such as draft approval time and the share of check-ins that lead to user action.
UX Principles
- Agents should only surface high-value information to avoid noise, using decision models to determine when to intervene in conversations.
- Review interfaces must be fast and integrated into tools like Slack, allowing users to provide feedback that feeds back into the agent's memory.
- Agents should operate where users already work, posting short, digestible updates and requests for review directly in communication channels.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.