Recently, I’ve tried quite a few AI Agents and looked at many projects about Agents. The more I see, the more I feel that what truly limits AI might no longer be model capabilities, but memory.

Most current Agents share a common problem: every time a new session starts, it’s almost like starting from scratch.

Bugs resolved yesterday, documents sorted, plans discussed, even your usage habits—none of it really stays. Many times, we’re not collaborating with AI, but repeatedly feeding it context.

That’s why, when I recently came across EverOS, I found this direction more interesting than simply competing on model parameters.

Instead of building yet another Agent, it builds a Memory OS behind the Agent—an infrastructure dedicated to managing long-term memory.

One thing I really appreciate is that it doesn’t treat Memory as a completely opaque black box.

EverOS stores all memories as Markdown, which can be viewed and edited locally, and version-controlled with Git. Under the hood, it uses Markdown + SQLite + LanceDB for retrieval and indexing, eliminating the need for additional complex components like MongoDB or Redis. For developers, when something goes wrong, you know where to look; if you want to modify something, you don’t have to guess what the model stored. This readable, controllable design feels more important than simply improving recall rate.

Also, it separates User Memory and Agent Memory into two independent growth paths—a pretty reasonable approach.

User information, preferences, and history make up one part; the experience, processes, and skills that the Agent summarizes during long-term use make up another. The two don’t mix. As usage increases, some repetitive tasks can gradually solidify into reusable Skills, instead of rewriting prompts every time.

Compared to many memory products that stay at the “store—retrieve—recall” level, EverOS draws my attention to its later concepts: Knowledge Wiki, Reflection, and Dreaming.

Simply put, it allows the Agent not only to remember what happened, but also to organize past knowledge into a continuously accumulating knowledge base, summarize experiences and extract patterns during idle time, and turn recurring problems into new capabilities. This approach is more like how humans learn, rather than just performing an information query.

I can’t say EverOS will definitely become the future standard, but at least it offers a direction I agree with: Memory shouldn’t just be a database, but rather the foundation for an Agent’s continuous growth.

In the future, whether it’s Claude Code, Codex, or various Coding Agents, Research Agents, Personal AIs—what truly determines the user experience ceiling may not be a few percentage points of model improvement, but who can truly own a set of transferable, accumulative, and evolvable long-term memory.

If you’ve also been following Agents, LLM applications, or AI Infra recently, I think this project is worth bookmarking.

⭐ GitHub:

I suggest starring it first, then when you have time, check out the README and overall architecture. At least among the open-source Memory projects I’ve seen recently, it’s one of the few that is both well-thought-out and close to real development scenarios.