Agentic AI

Current benchmarks rely on conversational recall, failing to measure if memory actually improves task execution for tool-using agents. The new MERIT benchmark addresses this by evaluating episodic tool-use tasks with explicit cost accounting and automated leak checks. It tracks token and dollar usage to determine the marginal utility of remembering facts across different difficulty levels.

• Replaces dialogue-centric benchmarks with tool-use task evaluation for agents
• Measures marginal utility of memory against explicit token and dollar costs
• Includes automated leak checks to verify true reliance on prior facts
• Tests updated-fact recall and controlled memory corruption scenarios

Amazon Bedrock AgentCore Memory now supports the IngestData API, allowing developers to push content directly into long-term memory without first creating a short-term memory event. This change decouples long-term memory extraction from the short-term memory pipeline, enabling independent adoption of persistent memory strategies. The API accepts both conversational payloads with user/assistant roles and raw JSON inputs.

• IngestData API bypasses short-term storage, reducing latency and storage overhead for persistent contexts.
• Supports conversational messages and JSON, offering flexibility in data ingestion formats.
• Developers can now implement long-term memory independently of short-term event workflows.
• Extracted records are immediately available via standard retrieval operations used by other memory types.

AutoFyn introduces a non-parametric training approach for long-horizon agents that updates persistent state rather than model weights. By leveraging a frozen base model, the system iteratively improves performance through verified reward signals distilled into memory files and repository state. An orchestrator manages specialized agents to explore solutions, while a verifier ensures task-grounded accuracy before updating the policy state for subsequent rounds.

• Avoids costly weight updates by updating persistent state like memory files instead.
• Uses a frozen base model to reduce computational overhead per iteration.
• Orchestrator explores multiple approaches while a verifier ensures accuracy.
• Rewards are distilled into durable state to guide future agent behavior.

SCAFFOLD introduces a framework for visual web agents that moves beyond flat skill caches by inducing parametric, executable skills from successful trajectories. It maintains a recursively composed hierarchy where higher-level skills invoke lower-level ones, allowing for principled compression of redundancy. This approach addresses the limitation of previous agents that discard procedural knowledge after isolated task execution.

• Moves beyond flat prompt-side caches to a parametric, executable skill library.
• Enables recursive composition where high-level skills invoke lower-level ones.
• Compresses redundancy in procedural knowledge across long-horizon web tasks.
• Agents retain and reuse knowledge instead of discarding it per task.

EdgeMem addresses the cost and data loss of compressing agent histories by preserving original interaction turns. It organizes these turns into a multi-anchor hypergraph using lightweight local processing based on content, temporal, and episodic cues. This structure allows retrieval engines to return direct source evidence to the LLM for final answer generation, rather than relying on pre-computed summaries.

• Eliminates compression costs by storing raw interaction turns instead of summaries.
• Uses a multi-anchor hypergraph to organize data via content, temporal, and episodic cues.
• Retrieval returns source evidence directly, reserving LLM usage for final answer synthesis.
• Reduces risk of discarding critical details before the specific query context is known.