LLMs are stateless. We built Dria Mem Agent to change that: Making memory a first-class feature. A 4B agent with local interoperable memory across Claude, ChatGPT and LM Studio. It turns LLMs from stateless chat into stateful agents with persistent human-readable memory.

Dria Mem Agent (4B) is nearly 60× smaller than Qwen3-235B-A22B-Thinking-2507 yet remains its closest rival, delivering a 35.7% performance boost over the base Qwen model with a score of 39.3%.

Mem-Agent is a memory proxy that runs locally and plugs into apps via MCP. It manages memory as markdown files in natural language, allowing you to: - read them like notes - edit them anytime - let the agent automatically retrieve and update them

It uses an Obsidian-style memory structure: - stores the profile and links - entities/ contains structured context - edits can be made manually or automatically A user-owned, interoperable memory layer.

Trained on three core subtasks: - Retrieve: gather relevant information, even across multiple hops. - Update: write new facts back into memory. - Clarify: ask questions when details are missing or conflicting. Most importantly, it supports natural language filters to hide or obfuscate sensitive information.

Ask Claude, “What did I tell ChatGPT about moving to London?” → mem-Agent pulls it from your ChatGPT history. Import your Notion workspace → Claude can now answer with company docs. Load your Google Docs folder → mem-Agent recalls information you need.

For coding agents, memory is an edge. Instead of relying on massive context windows and hitting rate limits: - Mem-Agent retrieves the right snippets from your local docs. - Condenses them and passes them to your agent. Smarter loops. Fewer tokens. Lower cost.

Built-in memory connectors → import/export context with ease. mem-agent-mcp repository supports: - ChatGPT exports - Notion - Nuclino - GitHub (live) - Google Docs (live)

Most “memory hacks” bolt RAG pipelines on top. Mem-Agent builds memory directly into the model. - Transparent: you see what’s retrieved, hidden, and updated - Portable: works across apps - Private: everything stays local

Mem-Agent is the quiet partner that remembers for you: - Keeps context alive - Reduces wasted tokens - Travels with you across apps 👉 Model: 👉 Blog: 👉 Repo:

If it is jot "inside" the LLM then the LLM is still stateless and you haven't changed anything

This looks very interesting!

this look very cool side note: I feel like I can’t take launch videos as seriously when the narration is obviously AI generated

This looks amazing. Nice work y'all!

Interesting approach to address the stateless nature of LLMs. Will be interesting to see how the human-readable memory feature scales.

cool

Memory , all you need is memory

beautiful video and website

Stateful agents are game-changing! Speaking of memory - Google Meet's new voice translation remembers conversation context across languages. Imagine agents that not only remember state but can communicate it across language barriers with preserved voice identity. That's the real unlock for global AI collaboration. Memory + Voice + Translation = True global intelligence 🧠

@MattVidPro This is cool, been meaning to build something similar! Will definitely check out your solution!

This is a great breakthrough! I had no idea projects like these were underway 🔥

Looks like a weekend project to me! Followed ✅

@SirMrMeowmeow

markdown memory is smart, you can read edit and trust the record later

This is a bandaid. You're still fundamentally limited by the context window.

Dria mem: memory first, scale

How does this compare to Letta?

How does the LLM actually ‘see’ and use this memory—are you piping it in at the prompt level, or is there a deeper integration layer?

What does it do? What’s the format of the memory saved? Is it some kinda database?

This was such a dope explainer

looks neat. should give it a shot

Wish it had a native Windows build too, but WSL works fine. Definitely see myself using this instead of the usual LLM maneuvering I have to do.

@grok your thoughts on this thread

Nice video! How did you make it? Any tools or template

Sounds cool but are there any production public project use cases?

this is exactly what's been missing from AI workflows, persistent memory changes everything for actual productivity

OMG!!!! You fixed it

why i can't join discord?

Precisely. We solve the review bottleneck by replacing subjective human audits with permanent Arweave blueprints.

@DhravyaShah

who edited the demo video??