A-MEM: Agentic Memory for LLM Agents

Abstract

Current memory systems for LLM agents rely on fixed structures that fail to capture the dynamic, evolving nature of knowledge. A-MEM addresses this by dynamically organizing memories in an agentic way, applying Zettelkasten principles to create interconnected knowledge networks through dynamic indexing and linking. Rather than passive storage, the system treats memory management as an active, agent-driven process. Published at NeurIPS 2025.

Key Contributions

• Structured memory creation: when new memories are added, the system generates comprehensive notes containing contextual descriptions, keywords, and tags
• Dynamic linking: the system analyzes historical memories to identify relevant connections, establishing links where meaningful similarities exist
• Memory evolution: new information can trigger updates to contextual representations and attributes of existing historical memories
• Empirical validation across six foundation models demonstrating improvements over existing approaches
• Open-source implementation (evaluation code and memory system) available on GitHub

Three-Part Memory Storage Architecture

1. Note Construction

When new information enters the system, the agent actively constructs a structured memory note containing:

• Core content — the raw information being memorized
• Contextual description — generated explanation of the content's significance
• Keywords and tags — extracted semantic labels for indexing and retrieval
• Metadata — temporal markers, source attribution, confidence levels

2. Link Generation

After note creation, the system performs retrospective analysis:

• Scans existing memory store for semantically related notes
• Evaluates multiple similarity dimensions (topical, temporal, causal)
• Establishes bidirectional links between related memories
• Creates emergent knowledge graphs from individual memory notes

3. Memory Evolution

The system actively maintains and updates its memory store:

• New information can modify contextual representations of existing memories
• Contradictory information triggers reconciliation processes
• Frequently accessed and linked memories gain higher retrieval priority
• Stale or superseded memories are marked but preserved for historical context

Zettelkasten Principles Applied

The system adapts the Zettelkasten method (slip-box note-taking) for AI agents:

• Atomicity — each memory note captures a single, self-contained concept
• Connectivity — value emerges from connections between notes, not individual notes
• Emergence — higher-level understanding develops organically from linked atomic notes
• Evolution — the knowledge network grows and reorganizes as new information arrives

Results

Evaluated across six foundation models on memory-intensive benchmarks. A-MEM consistently outperforms fixed-structure baselines (flat memory, hierarchical memory, summary-based memory) on tasks requiring:

• Multi-session context maintenance
• Knowledge integration across diverse sources
• Temporal reasoning over evolving information
• Complex query resolution requiring inference chains