README.md

EvolveMem: Self-Evolving Memory Architecture via AutoResearch

Extending SimpleMem with self-evolving retrieval infrastructure. The system autonomously researches its own architecture through LLM-driven closed-loop diagnosis.

Quick Start  ·  Architecture  ·  Results  ·  Evolution  ·  Citation

💡 Key Idea

Every existing memory system evolves what it stores but never how it retrieves. EvolveMem closes this gap.

The retrieval infrastructure (fusion weights, context budgets, answer styles, per-category overrides, ...) is exposed as a structured action space and optimized through an autonomous closed-loop:

This closed-loop self-evolution realizes an AutoResearch process: the system conducts the observe-hypothesize-experiment-validate cycle on its own architecture.

✨ Highlights

🚀 Quick Start

Installation

git clone https://github.com/aiming-lab/SimpleMem.git
cd SimpleMem/EvolveMem
pip install -r requirements.txt

Configuration

export OPENAI_API_KEY="your-key-here"
export OPENAI_API_BASE="https://api.openai.com/v1"   # or Azure endpoint
export LLM_MODEL="gpt-4o"

Run Self-Evolution

# Full evolution on LoCoMo (7 rounds)
python run_evolution.py --data data/locomo10.json --max-rounds 7

# Quick 3-round evolution
python run_evolution.py --data data/locomo10.json --max-rounds 3

# Start from pre-extracted memory cache
python run_evolution.py --use-cache cache.json --max-rounds 5

Run Benchmark Evaluation

# LoCoMo evaluation
python run_benchmark.py locomo --sample 0 --initial weak --max-rounds 3

# MemBench evaluation
python run_benchmark.py membench --agent FirstAgent \
    --categories simple comparative aggregative conditional \
    --initial weak --max-rounds 3

🏗️ Architecture

EvolveMem consists of three layers connected by a self-evolution feedback loop:

1. 🗄️ Structured Memory Store

2. 🔍 Multi-View Retrieval (Evolvable Action Space)

Fusion mode, per-view weights, context budgets, answer styles, and per-category overrides are all evolvable parameters.

3. 🧬 Self-Evolution Engine (AutoResearch)

The engine reads per-question failure logs, diagnoses root causes, and proposes targeted adjustments. Three safeguards ensure robustness:

Raw Conversations
  │
  ▼
┌─────────────────────────────┐
│   LLM-Based Extraction      │  ← Sliding window + retry + coverage verify
│   → Typed Memory Units      │
└─────────────┬───────────────┘
              ▼
┌─────────────────────────────┐
│   Multi-View Retrieval      │
│   BM25 ∪ Semantic ∪ Struct  │  ← Evolvable fusion (sum/weighted/RRF)
│   + Entity-swap             │
│   + Query decomposition     │
└─────────────┬───────────────┘
              ▼
┌─────────────────────────────┐
│   Answer Generation         │  ← Per-category style + verification
└─────────────┬───────────────┘
              ▼
┌─────────────────────────────┐
│   Evaluation + Diagnosis    │  ← LLM reads per-question failure logs
│   → Structured proposal     │
└─────────────┬───────────────┘
              ▼
┌─────────────────────────────┐
│   Meta-Analyzer             │  ← Revert / Explore / Apply
│   → Updated config θ        │
└─────────────┬───────────────┘
              │
              └──── Loop back to Retrieval ────┘

📊 Results

LoCoMo (Token-F1)

MemBench (Accuracy %)

🧬 Self-Evolution Trajectory

Starting from a minimal BM25-only baseline (F1 = 30.5%), the system autonomously discovers and activates retrieval mechanisms over 7 rounds:

Three configuration dimensions emerged from failure diagnosis that were not in the original design:

• 🔀 Query decomposition (splitting multi-hop questions into sub-queries)
• 🔄 Adversarial entity-swap (stripping misleading names before retrieval)
• ✅ Answer verification (second-pass LLM review of low-confidence outputs)

🔄 Cross-Benchmark Transfer

Continued evolution from a LoCoMo prior outperforms scratch evolution on MemBench (+16.6% relative) while also improving LoCoMo performance. Pareto improvement on both benchmarks.

📝 Citation

@article{evolvemem2026,
  title={EvolveMem: Self-Evolving Memory Architecture via AutoResearch for LLM Agents},
  author={Liu, Jiaqi and Ye, Xinyu and Xia, Peng and Zheng, Zeyu and Xie, Cihang and Ding, Mingyu and Yao, Huaxiu},
  journal={arXiv preprint arXiv:2605.13941},
  year={2026},
  url={https://arxiv.org/abs/2605.13941}
}