Resolution as Attention proposed to boost agent memory efficiency
Researchers introduce Resolution as Attention (RAA), a method that allocates perceptual detail to past interactions.
Source: Grafy Research LabResolution as Attention (RAA) is a new memory‑management concept for large‑language‑model (LLM) agents that treats the resolution of past interactions as a form of attention. The idea was presented in a paper posted to arXiv on 19 November 2026 by Siqi Xiang, Zhipeng Xu, Shengrui Liu, Zulong Chen, Chunyan Miao and Shijian Lu. According to the authors, modern LLM agents accumulate massive interaction histories when solving complex, long‑horizon problems, and existing memory‑management techniques focus largely on what to keep or retrieve. RAA adds a second dimension: how much perceptual detail each stored interaction should receive.
What is Resolution as Attention (RAA)?
RAA reframes memory allocation as an attention problem where the granularity of perception varies across the timeline of an agent’s experience. Instead of storing every past turn at full fidelity, the method assigns higher resolution (more tokens, richer embeddings) to recent or highly relevant events, while older or less critical steps are stored at lower resolution. The authors argue that this mirrors human recollection, where vivid details fade over time and only the gist remains. By coupling resolution with attention scores, an agent can dynamically balance memory size against the informational richness needed for a current decision.
Why does granularity matter for agent memory?
Long‑horizon tasks—such as multi‑step planning, open‑ended dialogue, or autonomous tool use—require agents to reference information that may have been generated many steps earlier. Keeping every token from every interaction quickly exceeds context windows and inflates inference cost. Prior approaches mitigate this by pruning irrelevant entries or by retrieving a fixed‑size subset, but they treat each retained entry as uniformly detailed. The authors note that this “one‑size‑fits‑all” view can waste compute on rarely used details and, conversely, starve the model of nuance when a seemingly minor past fact becomes crucial.
RAA’s variable‑resolution strategy directly addresses the trade‑off: it preserves the semantic outline of the entire history while allocating extra detail only where attention indicates it will most benefit the current query. In practice, this means an agent can maintain a longer effective horizon without exceeding its token budget.
How does RAA differ from prior memory‑management methods?
Existing techniques fall into two broad families:
- Retention‑oriented methods that decide which past interactions to keep (e.g., summarization, selective forgetting). These often rely on heuristics or learned importance scores but still store the chosen items at full resolution.
- Retrieval‑oriented methods that keep a large archive and use a similarity search to fetch a fixed number of entries at query time. Retrieval can be expensive and still returns uniformly detailed chunks.
RAA augments both families by introducing a resolution dimension. The paper describes a mechanism where attention weights are mapped to a resolution controller that determines token allocation for each entry before it is written to memory. Consequently, the memory buffer contains a heterogeneous mixture of high‑resolution and low‑resolution representations, all indexed by the same attention framework. The authors emphasize that this does not require extra model parameters; the resolution controller can be implemented as a lightweight mapping function on top of existing attention scores.
Potential impact on long‑horizon AI agents
If RAA can be integrated into the inference pipeline of LLM‑based agents, several practical benefits may emerge:
- Extended effective context windows – agents could reference events from hundreds of steps ago without hitting token limits, improving performance on tasks like multi‑day planning or complex codebase navigation.
- Reduced inference cost – by storing low‑resolution representations for older material, the total number of tokens processed per step drops, leading to faster response times and lower compute bills.
- More graceful degradation – when memory budgets are tight, the system can simply lower the resolution of the oldest entries, preserving overall task relevance rather than discarding whole chunks outright.
- Improved interpretability – the varying resolution offers a visual cue for developers to see which parts of an agent’s history the model deems important, aiding debugging and alignment work.
The authors caution that the approach still requires careful calibration of the resolution‑attention mapping to avoid under‑representing subtle but pivotal facts. They suggest future work on adaptive schedules that react to task‑specific signals.
Trying RAA with Graf by tools
While Grafy does not yet ship a built‑in RAA module, the research can be explored using existing Grafy apps:
- Reader – Load the arXiv paper directly into Grafy Reader (`/reader`) to experiment with question‑answering over the PDF. The tool’s ability to retrieve specific passages mirrors the retrieval aspect of RAA and lets you see how varying the amount of context changes answer quality.
- Chat – Use Grafy Chat (`/chat`) to prototype a simple agent that logs its dialogue turns. By manually down‑sampling older turns (e.g., summarizing them into a single sentence) you can simulate low‑resolution storage and observe the effect on later responses.
- Graph – The Graph app (`/graph`) lets you build a branching workflow where each node represents a memory entry at a chosen resolution. You can connect a resolution‑selection node to a retrieval node, effectively recreating the RAA pipeline without writing code.
- Code – For developers, Grafy Code (`/code`) provides an AI‑assisted coding environment where you can implement a lightweight resolution controller and plug it into an open‑source LLM agent framework (e.g., LangChain). The generated code can be run directly in the studio, giving you a hands‑on test of RAA’s trade‑offs.
By combining these apps, creators can prototype the core idea of RAA, compare it against baseline memory strategies, and gather empirical observations to inform future integration.
Frequently asked questions
How does RAA decide which past interactions receive higher resolution?
RAA uses the attention scores computed for the current query to infer relevance. Interactions with larger attention weights are assigned more tokens or richer embeddings, while those with lower weights receive a compressed representation. The mapping from score to resolution can be a simple linear function or a learned policy.
Can RAA be applied to multimodal agents that handle images or video?
The authors present RAA as a general attention‑based framework, so the principle extends to any modality that can be encoded as tokens. For visual data, lower‑resolution entries could correspond to down‑sampled feature maps or fewer image patches, preserving the same idea of allocating detail where attention indicates it is needed.
Does using RAA require changing the underlying LLM model?
No. RAA operates as a preprocessing and storage layer around the model. It leverages the model’s existing attention mechanism to guide resolution, so the base LLM architecture and weights remain unchanged.
Source: https://arxiv.org/abs/2611.24422
Source: Grafy Research Lab — Resolution as Attention for Efficient Agent Memory Perception


