MindLab-Research

MindLab-Research / delta-Mem

Public

Repo of Paper: delta-Mem: Efficient Online Memory for Large Language Models

19
0
100% credibility
Found May 18, 2026 at 18 stars -- GitGems finds repos before they trend. Get early access to the next one.
Sign Up Free
AI Analysis
Python
AI Summary

δ-mem (delta-Mem) is a research project that adds an efficient online memory system to large language models. Instead of treating each conversation as completely new, this system allows an AI to remember and build on previous interactions. When you chat with the AI, it writes important information into a compact memory space using a learning technique called delta-rule learning. Later, when you ask follow-up questions, the AI can retrieve and use this stored memory to provide more contextual responses. The project includes pre-trained memory adapters for popular AI models, an interactive chat demo, and evaluation benchmarks to measure memory performance. It supports three different memory writing strategies (TSW, SSW, MSW) and can be trained on custom data for specific applications.

How It Works

1
💭 You realize your AI forgets conversations

After chatting with an AI assistant for a while, you notice it can't remember things you discussed earlier in the conversation.

2
🧠 You discover a memory upgrade for AI models

You find a research project that adds persistent memory to AI assistants, so they can remember and build on past interactions.

3
📦 You download the memory adapter

You download a pre-trained memory module from the internet that's designed to work with your chosen AI model.

4
You choose how to use it
💬
Try the chat demo

Have a conversation with the AI and see it remember details from earlier messages.

📊
Run evaluation tests

Test how well the memory works on standard question-answering benchmarks.

5
🎓 You train your own memory (optional)

If you want custom memory for your specific use case, you can train the memory module on your own data.

✨ Your AI now has persistent memory

Your assistant can remember past conversations, answer questions about earlier topics, and maintain context over long interactions.

Sign up to see the full architecture

4 more

Sign Up Free

Star Growth

See how this repo grew from 18 to 19 stars Sign Up Free
Repurpose This Repo

Repurpose is a Pro feature

Generate ready-to-use prompts for X threads, LinkedIn posts, blog posts, YouTube scripts, and more -- with full repo context baked in.

Unlock Repurpose
AI-Generated Review

What is delta-Mem?

Delta-Mem adds a lightweight associative memory system to large language models. Instead of relying on full context retrieval or static parametric memory, it projects new information into a compact memory space and updates it dynamically during conversation using delta-rule learning. The system runs alongside a frozen attention backbone, letting models retain and recall information from extended interactions without endlessly growing context windows. Built in Python, it supports Qwen3 and SmolLM3 models with multiple write strategies for different memory granularities.

Why is it gaining traction?

The core appeal is solving context window fatigue without retrieval pipelines or fine-tuning for every new fact. Developers building long-running AI agents typically face a tradeoff between context length, retrieval complexity, and memory staleness. Delta-Mem sidesteps this by giving models an internal scratchpad that updates during inference. The released Qwen3-4B adapter on Hugging Face means you can experiment without training from scratch, and the bundled benchmarks against LoCoMo and MemoryAgentBench give concrete numbers rather than hand-wavy claims.

Who should use this?

AI engineers building stateful agents, chatbots that need to remember multi-session conversations, or anyone hitting context length limits with retrieval-augmented generation. Researchers evaluating memory mechanisms will appreciate the standardized eval suite. If you want plug-and-play memory for existing Qwen3 deployments, the adapter and demo scripts get you running quickly. Teams needing production-grade stability should treat this as experimental research code rather than a finished library.

Verdict

Delta-Mem addresses a real pain point with a clean architectural idea, but the 19 stars and 1.0% credibility score signal early-stage research software. The documentation is solid for a paper release, and the chat demo makes the concept tangible. Use it to explore whether associative memory fits your use case, but budget time for integration work before betting production on it.

Sign up to read the full AI review Sign Up Free

Similar repos coming soon.