RESEARCH

A-MEM

Zettelkasten-inspired self-organizing memory (NeurIPS 2025).

All researchBack to console

A-MEM: Zettelkasten-Inspired Self-Organizing Memory

A-MEM (Agentic Memory), presented at NeurIPS 2025, applies the Zettelkasten note-taking method to agent memory. The key insight: memories should be atomic, interconnected, and self-organizing.

Zettelkasten Principles Applied to Agents

The Zettelkasten method (invented by Niklas Luhmann) has three core principles that A-MEM adapts:

  1. Atomicity - Each memory captures exactly one idea or fact. No compound memories.
  2. Connectivity - Every memory links to related memories, forming a web of knowledge.
  3. Emergence - Structure emerges from connections, not from pre-defined categories.

Architecture

Loading diagram...

Key Operations

  • Split: Break multi-idea inputs into atomic memory units.
  • Link: Automatically discover and create connections between related memories.
  • Cluster: Emergent topic groupings form from link density (no manual tagging needed).
  • Retrieve: Follow links from seed memories to find relevant context.

Key Design Decisions

  • No pre-defined taxonomy: Categories emerge from connections, not from upfront schema design.
  • Bidirectional links: Every connection works both ways, enabling serendipitous discovery.
  • Self-organization: The memory system reorganizes itself as new memories arrive.
  • Retrieval via traversal: Start from a seed memory and follow links, rather than pure search.

Evaluation (NeurIPS 2025)

A-MEM showed improvements over flat memory stores on:

  • Multi-hop reasoning tasks (following connection chains)
  • Consistency (atomic memories reduce contradictions)
  • Scalability (link-based retrieval degrades more gracefully than brute-force search)

Relevance to Memory Platform

A-MEM's principles inform several potential enhancements:

  • Atomicity: Our memory creation guidance could encourage single-idea memories.
  • Linking: A memory_links table could capture relationships between memories.
  • Emergent structure: Tag clustering based on co-occurrence could surface natural groupings.
  • Traversal retrieval: "Find memories related to this memory" as a first-class operation.

The Zettelkasten approach is particularly relevant for long-running projects where the memory graph grows large and pre-defined categories become limiting.

References

  • A-MEM: Agentic Memory for LLM Agents (NeurIPS 2025)
  • Luhmann's Zettelkasten method and its digital adaptations
  • Graph-based memory organization for knowledge workers