EverMemOS

What is EverMemOS?

EverMemOS is an open-source memory operating system that gives AI agents persistent long-term memory, now developed under the name EverOS by the company EverMind. It records conversations, organises them into structured memories and retrieves relevant context when an agent needs it.

Large language models forget everything outside their context window. EverMemOS tackles this by extracting episodes, facts, user profiles and agent skills from interactions, storing them as Markdown files with local SQLite and LanceDB indexes, and serving them through a REST API. The January 2026 arXiv paper describes three stages: forming episodic traces, consolidating them into themes and profiles, and reconstructing context at query time.

EverMind describes itself as an AI infrastructure company focused on memory, and also lists a personal app called EverMe and an underlying layer called EverCore. The code is published on GitHub under the Apache 2.0 licence, and a managed EverOS Cloud service is offered for teams that prefer not to self-host.

Key Features

Human-readable storage is the feature that matters most, because every memory sits in an editable Markdown file you can audit, version and correct yourself.

• Structured long-term memory: EverMemOS stores user episodes and profiles alongside agent cases and skills. An agent can therefore recall what happened earlier, what a user prefers and which approaches worked before, across many sessions.
• Markdown source of truth: Memories are kept as readable Markdown files that you can edit, diff and version in Git. SQLite and LanceDB act as local indexes, so the files rather than a hidden database remain the canonical record.
• MemCells and MemScenes: The research paper describes turning dialogue into MemCells holding episodes, atomic facts and time-bound foresight signals. These are grouped into thematic MemScenes, which guide retrieval when the agent needs context later.
• Hybrid search API: A REST interface under /api/v2 exposes add, flush and search operations, with the older /api/v1 still supported. Search supports keyword and hybrid modes, so agents can combine exact terms with semantic matching.
• Model-agnostic design: The project FAQ says you can plug in OpenAI, Qwen, Llama or local models through an API wrapper. The self-hosted quick start asks for an OpenRouter key, which opens access to many hosted models.
• Managed cloud option: EverOS Cloud is offered as a fully managed service with automatic scaling and maintenance. Teams that do not want to run the Python server themselves can sign up at everos.evermind.ai instead.

Pros

Ownership is the strongest reason to pick EverMemOS: the Apache 2.0 code runs on your own hardware, with memories you can read and control directly.

• Free and open licence: The GitHub repository is released under Apache 2.0, which permits commercial use and modification. Startups can build memory into a product without licence fees or vendor lock-in.
• You own the data: Local-first storage keeps memories on your own machine or server as plain files. Privacy-conscious teams can inspect, back up and delete records without asking a third party.
• Published research: The design is documented in an arXiv paper from January 2026 with reported results on LoCoMo and LongMemEval. Engineers can read the method instead of trusting marketing claims alone.
• Light infrastructure: The current README lists no dependency on MongoDB, Elasticsearch, Milvus or Redis. A single Python 3.12 install with SQLite and LanceDB is enough for local development.
• Active community: The repository showed about 13,400 GitHub stars when we read it. That level of interest usually means more issues answered, examples shared and third-party integrations appearing over time.

Cons

Setup effort is the main hurdle, since EverMemOS is a developer framework that needs Python, a model key and integration work before it does anything useful.

• Built for developers: There is no point-and-click app for general users; setup means installing a Python package and running a server. Non-technical teams need an engineer to wire it into their agent.
• Name change confusion: EverMemOS has been renamed EverOS, while the paper and older articles still use the original name. Searching for documentation can lead you to outdated instructions or repository paths.
• Vendor-reported benchmarks: Accuracy figures such as 93.05 per cent on LoCoMo come from EverMind itself. Its own pages even quote two different HaluMem scores, so run your own evaluation before relying on them.
• Cloud pricing not settled: An EverMind blog post says the cloud is free during beta and lists $25 a month after launch. The homepage gives no price table, so long-term hosted costs are uncertain.
• LLM costs still apply: Memory extraction calls a language model, so you pay your model provider for each write and consolidation step. High-volume agents should estimate those token costs before going to production.

Who is Using EverMemOS?

Developers building AI agents are the primary audience, along with researchers studying long-term conversational memory.

• AI agent developers: Engineers building assistants and copilots use it as a memory backend. It lets their agents remember users and past tasks across separate sessions.
• Companion app builders: Teams making chat companions use profiles and episodes to keep personality and history consistent. This reduces repetitive questions and contradictory replies over time.
• Researchers: Academics compare it with other memory systems on long-conversation benchmarks. The open code and paper make results easier to reproduce and extend.
• Privacy-focused teams: Organisations with strict data rules self-host it on their own servers. Markdown storage also makes audits and deletion requests easier to handle.

Pricing

The open-source code is free under Apache 2.0, and EverOS Cloud is reported as free in beta with a planned paid tier (EverMind site and blog, checked October 2026).

We found no public price table on the EverMind homepage, and its pricing URL returned an error, so the cloud figure rests on one company blog post.

Disclaimer: hosted pricing for beta products often changes at launch, so confirm current terms in the EverOS Cloud console.

What Makes EverMemOS Unique?

Plain Markdown as the canonical memory store is the clearest difference from rivals such as Mem0 or Zep, which centre on database or graph storage. You can open a memory, fix a wrong fact by hand and commit the change to Git like any other file.

The second difference is its research grounding. The MemCell and MemScene design, with time-bound foresight signals and evolving user profiles, is set out in a public paper and benchmarked on LoCoMo, LongMemEval and PersonaMem v2. Few memory layers document their method in that much detail.

Compatibilities and Integrations

An HTTP API is the main integration point, so EverMemOS works with any agent stack that can send web requests, plus Python and MCP clients.

• REST API: Agents call add, flush and search endpoints under /api/v2 over HTTP. Any language that can make web requests can use EverMemOS as a memory service.
• Python and MCP: A Python package covers local use, and the README mentions Model Context Protocol support. MCP lets compatible assistants reach memory without custom glue code.
• Agent frameworks: The project FAQ says it can act as a plug-in memory backend for LangGraph and Haystack. That lets existing pipelines add long-term memory without a rewrite.
• Models and storage: OpenAI, Qwen, Llama and local models are listed as supported, with OpenRouter used in the quick start. Data lives in Markdown with SQLite and LanceDB indexes.

EverMemOS Tutorials

A local test install takes under an hour for a developer comfortable with Python and API keys.

1. Read the EverOS README on GitHub (EverMind-AI/EverOS) and check you have Python 3.12 or later.
2. Install the package with pip or uv in a fresh virtual environment.
3. Add your OpenRouter key, or point the configuration at another model provider you use.
4. Start the local server with the everos server start command and confirm the API responds.
5. Send a test conversation to the add endpoint, then query it with the search endpoint.
6. Open the generated Markdown files to inspect what was stored, and edit or delete records as needed.

How We Rated It

EverMemOS scores 4.0 out of 5 for us, because free open-source code and local ownership outweigh a developer-only setup and vendor-reported benchmarks.

Summary

EverMemOS, now called EverOS, is a strong pick for developers who want long-term agent memory they can self-host, inspect and control. Its Apache 2.0 licence and Markdown storage suit privacy-conscious teams and researchers. The main limitation is that it is a developer framework, not a ready-made app, and its benchmark numbers come from the vendor. Hosted pricing also looks unsettled while the cloud remains in beta. Start with a local install on a small test agent and measure recall on your own conversations before you commit.