GAM: Hierarchical Graph-based Agentic Memory for LLM Agents (April 2026)

Free

Hierarchical graph-structured memory with role-aware modulation and temporal/confidence weighting; training-free, evaluated across multiple model scales

FreeFree tier

Type

Open Source

Visit Website GitHub (40 stars)

About GAM: Hierarchical Graph-based Agentic Memory for LLM Agents (April 2026)

GAM is a hierarchical graph-based agentic memory framework for Large Language Model (LLM) agents, designed to sustain coherent long-term interactions by decoupling memory encoding from consolidation. It uses an event progression graph to capture ongoing dialogue and integrates information into a topic associative network only upon semantic shifts, minimizing interference while preserving long-term consistency. The framework also employs a graph-guided, multi-factor retrieval strategy to enhance context precision. Evaluated on the LoCoMo and LongDialQA benchmarks, GAM consistently outperforms state-of-the-art baselines in both reasoning accuracy and efficiency.

Key Features

Decouples memory encoding from consolidation to resolve tension between rapid context perception and stable knowledge retention

Uses an event progression graph to isolate ongoing dialogue and minimize interference from transient noise

Integrates into a topic associative network only upon semantic shifts to preserve long-term consistency

Graph-guided, multi-factor retrieval strategy enhances context precision

Training-free, directly applicable to existing LLM agents

Pros & Cons

Pros

Outperforms state-of-the-art baselines on LoCoMo and LongDialQA benchmarks
Explicitly addresses the conflict between context perception and stable knowledge retention
Preserves long-term consistency by minimizing interference from transient noise
Graph-guided retrieval improves context precision without additional training

Best For

Sustaining coherent long-term interactions in conversational AI and LLM agentsManaging evolving narratives in dialogue systemsImproving reasoning accuracy and efficiency in long-context tasksEnhancing memory retention for agents operating over extended conversations

FAQ

What problem does GAM solve?

GAM addresses the tension between acquiring new information and retaining prior knowledge in LLM agents, enabling coherent long-term interactions without interference from transient noise.

How does GAM work?

GAM explicitly decouples memory encoding from consolidation by isolating ongoing dialogue in an event progression graph and integrating it into a topic associative network only upon semantic shifts. It also uses a graph-guided multi-factor retrieval strategy for enhanced context precision.

On which benchmarks was GAM evaluated?

GAM was evaluated on LoCoMo and LongDialQA benchmarks, consistently outperforming state-of-the-art baselines in both reasoning accuracy and efficiency.