Facts as First Class Objects: Knowledge Objects for Persistent LLM Memory

2026/03/18 by Oliver Zahn, Simran Chana · 1 voice
Computer Science · #Advanced Graph Neural Networks #Adversarial system #Artificial neural network #Benchmark (surveying) #Big Data and Digital Economy #Class (philosophy) #Context (archaeology) #Embedding #Memory model #Mnemonic #Recall #Recurrent neural network #Topic Modeling #cs.AI

paper · pdf · doi:10.48550/arxiv.2603.17781

openalex publication_date 2026/03/18 · arxiv published 2026/03/18 · arxiv updated 2026/03/18 · openalex created_date 2026/03/20 · openalex updated_date 2026/07/28

Abstract

Large language models increasingly serve as persistent knowledge workers, with in-context memory - facts stored in the prompt - as the default strategy. We benchmark in-context memory against Knowledge Objects (KOs), discrete hash-addressed tuples with O(1) retrieval. Within the context window, Claude Sonnet 4.5 achieves 100% exact-match accuracy from 10 to 7,000 facts (97.5% of its 200K window). However, production deployment reveals three failure modes: capacity limits (prompts overflow at 8,000 facts), compaction loss (summarization destroys 60% of facts), and goal drift (cascading compaction erodes 54% of project constraints while the model continues with full confidence). KOs achieve 100% accuracy across all conditions at 252x lower cost. On multi-hop reasoning, KOs reach 78.9% versus 31.6% for in-context. Cross-model replication across four frontier models confirms compaction loss is architectural, not model-specific. We additionally show that embedding retrieval fails on adversarial facts (20% precision at 1) and that neural memory (Titans) stores facts but fails to retrieve them on demand. We introduce density-adaptive retrieval as a switching mechanism and release the benchmark suite.

Citations

• Titans: Learning to Memorize at Test Time
• Leave No Context Behind: Efficient Infinite Context Transformers with Infini-attention
• Fine-Tuning LLaMA for Multi-Stage Text Retrieval
• MemGPT: Towards LLMs as Operating Systems
• RECOMP: Improving Retrieval-Augmented LMs with Compression and Selective Augmentation
• Ring Attention with Blockwise Transformers for Near-Infinite Context
• Lost in the Middle: How Language Models Use Long Contexts
• FlashAttention: Fast and Memory-Efficient Exact Attention with IO-Awareness
• Memorizing Transformers
• Unsupervised Dense Information Retrieval with Contrastive Learning
• Salient Phrase Aware Dense Retrieval: Can a Dense Retriever Imitate a Sparse One?
• RocketQA: An Optimized Training Approach to Dense Passage Retrieval for Open-Domain Question Answering
• Big Bird: Transformers for Longer Sequences
• Approximate Nearest Neighbor Negative Contrastive Learning for Dense Text Retrieval
• Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks
• Dense Passage Retrieval for Open-Domain Question Answering
• Longformer: The Long-Document Transformer
• Compressive Transformers for Long-Range Sequence Modelling
• RotatE: Knowledge Graph Embedding by Relational Rotation in Complex Space

Discussions

• “Facts as First Class Objects” proposes Knowledge Objects (KOs) for persistent LLM memory, and benchmarks them against in-context memory, embedding retrieval, and neural memory. KOs sharply improve fa [bsky, 1 points, 1 comments]

Related

• MeMo: Memory as a Model
• LLMs as Repositories of Factual Knowledge: Limitations and Solutions
• Do Large Language Models Know about Facts?
• Are LLMs Really Not Knowledgeable? Mining the Submerged Knowledge in LLMs' Memory
• Pre-training Limited Memory Language Models with Internal and External Knowledge