The Memory Control Loop

For long-horizon tasks, AI agents often suffer from context rot—forgetting previous steps or contradicting themselves. Rather than treating memory as a static database, view it as a write-manage-read control loop. This harness acts as an external cognitive layer for models that lack durable memory.

Harness Architecture & Recall Policies

The harness consists of three primary blocks:

• Core: A persistent trace of the agent's actions always visible to the model.
• Archival: A long-term storage mechanism for information across sessions.
• Recall: The decision-making layer that retrieves relevant data.

Testing various recall strategies reveals that a ranked policy—which prioritizes decisions made during the agent's execution—outperforms simple vector-based RAG. Even when provided with an 'oracle' (ground truth memory), models do not always achieve 100% accuracy, suggesting that the model's ability to interpret and utilize retrieved data is as critical as the retrieval mechanism itself.

Performance and Cost Trade-offs

• Context Fit: If a task fits entirely within the model's context window, adding a memory harness provides no performance benefit and only increases latency and cost.
• Long-Horizon Tasks: For tasks exceeding the context window, a structured memory harness is essential.
• Efficiency: A well-designed recall policy reduces token usage by preventing the agent from pursuing incorrect paths, effectively lowering operational costs compared to 'dumb' retrieval or no-memory approaches.

Practical Implementation

Experiments conducted on local hardware (M3 Ultra, 96GB RAM) using Qwen 2.5 72B and DeepSeek V4 Flash demonstrate that local models are increasingly viable for agentic tasks. While local execution requires serial processing (limiting throughput), it offers total sovereignty over the data pipeline, allowing for granular control over evaluation traces and memory management.