FluxMem: Rethinking Memory as Continuously Evolving Connectivity

A compelling new paper on arxiv (2605.28773) introduces FluxMem, a memory architecture for LLM agents that treats memory not as static storage but as a dynamically editable heterogeneous graph that progressively refines its own topology.

The authors identify two key failure modes in existing memory systems: inaccurate memory connections (both under-connection and over-connection) and inflexible memory unit content that cannot adapt to new contexts.

Architecture: Three-Layer Memory Graph

FluxMem organizes memory into three interconnected layers:

Layer	Content
Semantic	Static factual knowledge
Episodic	State-action trajectories from experience
Procedural	Distilled reasoning templates and skills

Three-Stage Pipeline

The system evolves through a structured pipeline:

1. Initial Connection Formation — Uses hybrid retrieval combining dense embeddings, BM25, and LLM-based matching to create initial graph edges.
2. Feedback-Driven Refinement — Three operations refine the graph: link expansion (adding missed connections), link pruning (removing spurious connections), and content reshaping (updating node content).
3. Long-Term Consolidation — Episodic clustering groups related experiences, and PEMS-guided skill induction distills reusable procedural knowledge.

Results

FluxMem achieves state-of-the-art results across multiple benchmarks:

• LoCoMo: 95.06 (using GPT-4.1-mini)
• Mind2Web: SR 8.1
• GAIA: 64.85% (K2 backbone) — a 12.73% absolute improvement over the baseline

The code is available at github.com/zjunlp/LightMem.

---

Source: arXiv 2605.28773 — FluxMem