AI Agents Recall Facts When Asked Directly but Fail to Apply Them Unprompted

·1 views

A study published on arXiv on 27 July 2026 by Ruizhe Li and colleagues tested six memory systems across 125 tasks, finding that AI agents answered direct questions naming a stored fact correctly 76–100% of the time. However, when tasks implicitly depended on the same fact without naming it, end-to-end success rates collapsed to at most 14.4% across all six systems. Manually placing the fact into the model's context raised indirect task performance to 84%, leading the authors to conclude the gap reflects an access problem, not a capability one. The core issue, which the researchers term 'open problem routing', is that current systems lack reliable mechanisms to surface relevant facts before a query arrives. The findings suggest that AI memory systems can store and retrieve information on demand while still failing to use that information when it matters most.

Read the full story at DEV Community

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

Insufficient content to generate a reliable headline

The source provided does not contain enough article text to produce an accurate, fact-based summary. Only metadata such as the article URL, Hacker News comments link, and engagement stats were included. No substantive content about family preparedness was available for review. To ensure accuracy, please provide the full article text for summarization.

0 comments Read more at Hacker News

ProgrammingDEV Community ·

How One AI Pipeline Learned to Catch Valid but Wrong-Shaped JSON Responses

On March 3, 2026, a content automation pipeline built by an IT analyst failed not because an AI model returned invalid JSON, but because it returned a bare array instead of the expected envelope object. The error passed JSON parsing cleanly but was caught only later by Zod schema validation, highlighting a lesser-known second layer of structured-output failure. To fix the issue, the developer embedded the exact JSON schema directly into the prompt and added a deterministic code fallback to wrap bare arrays automatically. The approach was later codified as a standing rule across two levels of project configuration to ensure it applied to all future work. The lesson proved broadly relevant when the same pipeline was extended to support three interchangeable AI backends, confirming that prompt-plus-code defenses are necessary regardless of the underlying model.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Agent Memory Explained: Four Types Every AI Developer Should Know

AI agent memory is the mechanism that allows agents to retain information — such as facts, preferences, and past actions — across sessions and tool handoffs, rather than resetting with each new conversation. Developer Anthony Conti of Astra AI outlines four distinct memory types: working memory (context window), long-term semantic memory (external storage), episodic memory (timestamped event logs), and procedural memory (learned behavioral patterns). As agents increasingly handle multi-day tasks in 2026, the limitations of even million-token context windows have made persistent external memory a critical engineering concern. Most practical systems require only working memory combined with long-term semantic memory, while episodic and procedural types become essential for complex, extended workflows. Conti also clarifies that retrieval-augmented generation is a retrieval technique, not a memory system itself — agent memory additionally governs what information gets written and when.

0 comments Read more at DEV Community

ProgrammingHacker News ·

LibreOffice hits download record after proudly announcing no AI features

LibreOffice, the open-source office suite, has broken its own download records following a public declaration that it contains no AI features. The announcement appears to have resonated strongly with users who are increasingly wary of AI being embedded in productivity software. The surge in downloads suggests a growing segment of users actively seeking software free from AI integration. This trend highlights a counter-movement to the industry-wide push by major software companies to incorporate AI into their products.

0 comments Read more at Hacker News