Memora — Adaptive AI Agents with Memory (No Fine-Tuning Required)

September 3, 2025

This content originally appeared on DEV Community and was authored by Danda Pranav Krishna

The Problem with Today’s AI Agents

Large Language Models (LLMs) are powerful, but they don’t truly learn from experience.
Each interaction is isolated — no memory of past attempts, no cumulative knowledge.

Fine-tuning can help, but it’s:

Expensive
Rigid
Slow to iterate

If we want truly adaptive agents, we need a better paradigm.

The Idea: Memory-Augmented Reinforcement Learning

Instead of retraining the model itself, Memora introduces memory into the loop.

Episodic Memory — stores past experiences (success + failure).

Case Retrieval — brings up the most relevant past examples for new tasks.

Memory Rewriting — updates knowledge dynamically with feedback.

This shifts the agent’s learning from parameter updates → to retrieval + reasoning.

How Memora Works

The architecture follows a Planner–Executor cycle:

Meta-Planner (System 2):

Strategically breaks down complex problems.

Leverages memory for analogical reasoning.

Executor (System 1):

Executes steps sequentially.

Writes results back to memory.

This means the agent improves with experience without touching the base model weights.

Key Results

GAIA benchmark: 87.88% validation (outperforming GPT-4 baseline).

DeepResearcher benchmark: +4.7–9.6% gain on out-of-domain tasks.

Local LLMs (Qwen2.5-14B, LLaMA): achieved near GPT-4 performance — on a consumer MacBook (M4).

Why This Matters

Continual learning without retraining.

Cost efficiency — runs on everyday hardware.

Interpretability — every decision can be traced back to memory.

Scalability — agents adapt in real time.

Try It Yourself

# Install Ollama
curl -fsSL https://ollama.ai/install.sh | sh

# Pull a local model
ollama pull qwen2.5:14b

# Clone the repo
git clone https://github.com/Agent-on-the-Fly/Memora
cd Memora && pip install -r requirements.txt

# Run the agent
python client/agent.py

GitHub Repo