Agent harnesses have emerged as the primary framework for developing intelligent systems. These structures define how agents interact with tools and data, and they are deeply intertwined with memory management. The choice of harness determines not only the technical capabilities of an agent but also the control users retain over its memory.
Agent Harnesses as the Foundation
The evolution of agentic systems has seen a shift from simple retrieval-augmented generation (RAG) chains to more complex architectures. Early tools like LangChain and LangGraph laid the groundwork, but agent harnesses have become the standard for building agents that can perform tasks requiring sequential reasoning and external data access.
Examples of harnesses include Claude Code, Deep Agents, Pi (used by OpenClaw), Codex, Letta Code, and others. These frameworks provide the scaffolding needed to integrate large language models with tools, enabling agents to process information beyond their training data.
Memory and Its Integration with Harnesses
Memory is a critical component of agent functionality, enabling systems to retain context across interactions and build personalized experiences. However, the relationship between memory and harnesses is complex. As Sarah Wooders explains, memory is not a separate service but a core responsibility of the harness itself.
The harness manages both short-term and long-term memory. Short-term memory includes conversation history and tool responses, while long-term memory requires persistent storage and updates. The harness determines how context is maintained, how metadata is structured, and how interactions are stored for future reference.
Ownership of Memory and the Risks of Closed Systems
Using a closed harness—especially one behind a proprietary API—can lead to loss of control over memory. For instance, stateful APIs like OpenAI’s Responses API or Anthropic’s compaction system store data on their servers, limiting users’ ability to transfer or manage it independently.
Closed systems also introduce uncertainty. Tools like Claude Code, which is not open-source, obscure how memory is managed internally. This lack of transparency makes it difficult to migrate between platforms or understand the structure of stored data.
The Future of Memory in AI Agents
As agents become more sophisticated, memory will play an increasingly vital role. It enables systems to learn from user interactions, creating a feedback loop that improves performance over time. However, this dependency on memory also raises concerns about lock-in.
Model providers are incentivized to centralize memory within their platforms. Anthropic’s Claude Managed Agents exemplify this trend, locking all functionality behind APIs. Even open-source tools like Codex generate encrypted summaries that are tied to specific ecosystems, limiting cross-platform compatibility.
Open Memory, Open Harnesses
The long-term viability of agent systems depends on open standards for memory and harnesses. Closed systems risk creating monopolies where users lose access to their data once they switch providers. This is a stark contrast to the relatively easy transition between model providers in the past, which was possible due to stateless APIs.
A real-world example highlights this issue: an internal email assistant built on OpenClaw’s framework lost its memory after accidental deletion. Rebuilding it from scratch required retraining all user preferences, underscoring how memory enhances usability and personalization.
The industry is still in the early stages of defining memory systems. While long-term memory may not be a priority for many MVPs, the potential for proprietary datasets remains significant. As best practices emerge, the line between open and closed systems will likely blur, but for now, transparency and control remain essential.
MT Labs helps companies across Singapore deploy AI tools they actually own. Private infrastructure, no recurring cloud subscriptions, and a setup built around how your team already works. Whether you’re exploring your first AI use case or consolidating scattered tools into one system, we’ll walk you through it. Get in touch and let’s figure out what makes sense for your business.
