AI Research

Hierarchical Graph Memory for LLM Agents with Path-... | AI Research

Key Takeaways

HiGram is a framework designed to improve how LLM agents manage long-term memory.
By organizing information into a hierarchical graph rather than a flat stru...
Agents for long term reasoning require a memory that can be efficiently and effectively updated over time, as new facts and external feedback continue to arrive.
Recently, graph memory has been adopted to offer structural organization for multi-hop retrieval and reasoning.
However, existing methods store all memories in a flat graph, and accumulated historical memories can introduce irrelevant contexts and increase the cost of evidence selection during retrieval.

Paper AbstractExpand

Agents for long term reasoning require a memory that can be efficiently and effectively updated over time, as new facts and external feedback continue to arrive. Recently, graph memory has been adopted to offer structural organization for multi-hop retrieval and reasoning. However, existing methods store all memories in a flat graph, and accumulated historical memories can introduce irrelevant contexts and increase the cost of evidence selection during retrieval. Moreover, they typically update memory units independently, requiring repeated unit-wise rewrite to cover related changes. To address these issues, we propose HiGram, an evolving hierarchical graph memory framework with path-level localization and rewriting. Specifically, we first propose a hierarchical graph memory, which organizes the memory into coarse-to-fine architecture composed of upper-level nodes and MemoryUnits, thereby reducing the amount of irrelevant information during retrieval. We further propose MicroGraph-based path-level localization, which leverages query and update conditioned MicroGraphs to identify support subgraph and evidence path before rewrite. Finally, we propose a coordinated rewriting method that jointly revises intra-unit memory and inter-unit dependencies, enable valid dependency structures updating in the localized evidence path. Experiments on benchmarks for long-term conversational question answering and conflict-aware memory evaluation demonstrate that our method demonstrate substantial improvements over baselines in answer quality and token efficiency. Besides, our method improves answer accuracy and query-valid evidence selection under dynamic, static, and conditional conflicts.

HiGram is a framework designed to improve how LLM agents manage long-term memory. By organizing information into a hierarchical graph rather than a flat structure, the system reduces the amount of irrelevant data retrieved during reasoning and enables more efficient updates when new facts or feedback arrive.

Addressing Memory Bloat

Existing graph-based memory systems often store all historical information in a single, flat graph. As these memories accumulate, retrieving relevant evidence becomes costly and prone to including irrelevant context. Furthermore, these systems typically update memory units in isolation, which can lead to outdated dependencies and the need for repeated, inefficient rewrites. HiGram addresses these issues by separating memory into a coarse-to-fine architecture, where upper-level nodes provide abstract organization and MemoryUnits store specific factual details.

Path-Level Localization

To improve the precision of memory updates, HiGram uses a process called MicroGraph-based path-level localization. When a query and an update are received, the system constructs temporary memory units to identify relevant "anchors." These anchors help the system select a specific, localized support subgraph from the global memory. Within this subgraph, the framework identifies the specific "evidence path"—a chain of connected memory units and dependencies—that is relevant to the current query. This narrows the scope of the update, ensuring that only the necessary parts of the memory are modified.

Coordinated Rewriting

Once the relevant evidence path is localized, HiGram performs a coordinated rewrite. This process involves two simultaneous actions:

Intra-unit rewriting: Updating the internal states of the affected memory units based on new information.
Inter-unit rewriting: Revising the dependency structures that connect these units to ensure the entire evidence chain remains consistent.
By updating both the facts and their relationships within a bounded region, the system avoids the need for global searches or repeated, unit-by-unit modifications.

Performance and Evaluation

The researchers evaluated HiGram using the LoCoMo benchmark for long-term conversational question answering and the MemConflict benchmark for conflict-aware memory evaluation. According to the study, HiGram achieved improvements in answer quality and token efficiency compared to existing baselines. The results indicate that the framework also improved answer accuracy and the selection of valid evidence when tested against dynamic, static, and conditional conflicts.

Comments (0)

No comments yet

Be the first to share your thoughts!