Agent Context Management

Agent context management refers to the internal mechanisms an agent-architecture uses to maintain state, memory, and experience over time. Effective context management allows an agent to learn from past interactions, perform long-horizon reasoning, and personalize its behavior. Modern large language models (LLMs) increasingly depend on efficient long-context processing and generation mechanisms, including sparse attention, retrieval-augmented generation (RAG), and compressed contextual memory, to support complex reasoning.

Memory Architectures and State Management

General Principles and Standards

memorywire is a vendor-neutral JSON-Schema wire format for agent memory operations, defining five operations (remember, recall, forget, merge, expire) over four memory types (semantic, episodic, procedural, emotional) memorywire: A Vendor-Neutral Wire Format for Agent Memory Operations.

Cognitive and Evolving Memory

SaliMory is a framework for training a single model to manage a cognitively-structured memory of user facts, preferences, and working memory SaliMory: Orchestrating Cognitive Memory for Conversational Agents.
It uses hierarchical stage-wise process rewards and reward-decomposed contrastive refinement to provide supervision for memory operations like filtering, consolidation, and recall SaliMory: Orchestrating Cognitive Memory for Conversational Agents.
Palimpsa is a self-attention model that treats In-Context Learning (ICL) as a continual learning problem, using Bayesian metaplasticity to manage the stability-plasticity dilemma and expand memory capacity Learning to Remember, Learn, and Forget in Attention-Based Models.
The LifeSkill framework uses Online Skill Internalization to continuously improve a policy model during test-time interaction by converting trajectories into reward signals, which internalizes reasoning capabilities into parameters and avoids context bloat Learning While Acting: A Skill-Enhanced Test-Time Co-Evolution Framework for Online Lifelong Learning Agents.

Dynamic and State-Grounded Context

State-Grounded Dynamic Retrieval (SGDR) is an online skill learning method for web agents that retrieves skills based on both the task goal and the current webpage state Online Skill Learning for Web Agents via State-Grounded Dynamic Retrieval.
The AgentMob framework uses adaptive evidence-controlled decision making, triggering iterative tool use for ambiguous cases while using a fast path for routine ones Towards Efficient and Evidence-grounded Mobility Prediction with LLM-Driven Agent.
The Ptah multi-agent harness uses a Visual Working Memory to maintain source-aligned images for generating interleaved multimodal reports Towards Verifiable Multimodal Deep Research: A Multi-Agent Harness for Interleaved Report Generation.

Retrieval-Augmented Context

Retrieval and Reranking Strategies

The R3G framework for VQA uses a Reasoning-Retrieval-Reranking pipeline, where a reasoning plan first specifies required visual cues before a two-stage retrieval and reranking process selects evidence images R3G: A Reasoning-Retrieval-Reranking Framework for Vision-Centric Answer Generation.
Deontic Agentic Reasoning (DAR) allows a model to interact with statutes on demand, retrieving relevant rules as needed during the reasoning process DAR: Deontic Reasoning with Agentic Harnesses.
The BioManus agent plans over a structured capability graph of bioinformatics tools, retrieving compact task-specific subgraphs to synthesize workflow scaffolds, decoupling planning complexity from the total tool inventory size Beyond Prompt-Based Planning: MCP-Native Graph Planning-based Biomedical Agent System.

Representation and Indexing

Hyperbolic Retrieval-Augmented Generation (HyRAG) models tree-like external knowledge bases in hyperbolic space to better retain hierarchical structures and improve semantic granularity in retrieval Generalizing Graph Foundation Models via Hyperbolic Retrieval-Augmented Generation.
Single-stage Sparse Retrieval (SSR) uses a Sparse Autoencoder (SAE) to project token embeddings into a high-dimensional sparse representation, replacing clustering-based indexing with an inverted index for faster multi-vector retrieval No More K-means: Single-Stage Sparse Coding for Efficient Multi-Vector Retrieval.
MM-BizRAG uses a document structure-aware split to apply layout-specific parsing to enterprise documents and decouples retrieval representations from generation context to improve multimodal Q&A MM-BizRAG: Rethinking Multimodal Retrieval-Augmented Generation for General Purpose Enterprise Q&A.

Context Compaction and Efficiency

Compression and Summarization

The Upfront CoT (UCoT) framework uses a post-reasoning paradigm where a lightweight "compressor" model generates a compressed Chain-of-Thought as soft tokens, which are then used as context by a larger "executor" LLM Can Reasoning Path still be Effective as Input? Bridging Post-Reasoning to Chain-of-Thought Compression.

Bounded Context and KV Cache

The VulnAgent-R2 auditing framework uses bounded context optimization as one of its modules to manage evidence gathering VulnAgent-R2: Evidence-Calibrated Multi-Agent Auditing for Repository-Level Vulnerability Detection.

Evaluation and Failure Modes

Benchmarks and Challenges

MemoryDocDataSet is a benchmark designed to evaluate an agent's ability to jointly use multi-session conversational memory and perform reasoning over long documents MemoryDocDataSet: A Benchmark for Joint Conversational Memory and Long Document Reasoning.
It reveals a "joint-retrieval gap," where systems proficient in document-only retrieval see performance collapse when they must first use conversational history to identify the correct document MemoryDocDataSet: A Benchmark for Joint Conversational Memory and Long Document Reasoning.

Memory Security and Robustness

Persistent memory in agents introduces the risk of memory poisoning, where a single adversarial write can have long-term influence on behavior From Untrusted Input to Trusted Memory: A Systematic Study of Memory Poisoning Attacks in LLM Agents.
A systematic study identified four memory write channels and nine structural vulnerabilities in agent systems that enable six classes of memory poisoning attacks From Untrusted Input to Trusted Memory: A Systematic Study of Memory Poisoning Attacks in LLM Agents.
The MPBench benchmark was developed to evaluate these attacks, showing that agents designed for more aggressive memory writing and retrieval are more exploitable From Untrusted Input to Trusted Memory: A Systematic Study of Memory Poisoning Attacks in LLM Agents.