EverMind-AI/EverOS

EverOS – Persistent, Local‑First Memory for AI Agents

What it is – EverOS is a Python package that gives any LLM‑powered agent a durable, searchable memory store that lives on the user’s machine. It records conversations, files, and “agent trajectories” as plain‑text Markdown files, then keeps a lightweight SQLite index and a LanceDB vector index for fast retrieval. The design is deliberately local‑first: no external databases, cloud services, or managed vector stores are required unless you choose to add them.

Why it matters – Most agent‑memory libraries keep state in opaque databases or remote vector stores, making the data hard to inspect, version‑control, or edit by hand. EverOS treats the Markdown files as the single source of truth – they are human‑readable, Git‑trackable, and can be edited directly. A file‑watcher syncs any changes to the SQLite/LanceDB indexes automatically, so the memory stays consistent without extra tooling.

Core concepts

Concept	How EverOS handles it
Storage format	Human‑readable *.md files (canonical source)
Indexes	SQLite for metadata + LanceDB for vector embeddings
User vs. Agent tracks	Separate “episodes/profile” for users and “cases/skills” for agents
Retrieval scopes	Search can be filtered by user_id, agent_id, app_id, project_id, session_id
Knowledge Wiki	Editable Markdown knowledge pages with taxonomy and CRUD APIs
Reflection	Offline process that clusters episodes and refines profiles/skills across sessions

Getting started (quick‑run)

1. Install – pip install everos (or uv pip install everos).
2. Demo – everos demo runs a self‑contained example that shows ingest → index → recall, no API key needed.
3. Configure LLM – Run everos init to create ~/.everos/everos.toml and add a single OpenRouter API key (e.g., openai/gpt-4.1-mini).
4. Start the server – everos server start launches a local HTTP API (default http://127.0.0.1:8000).
5. Add memory – POST JSON to /api/v2/memory/add with a session, app, project, and a list of messages.
6. Flush – Signal end‑of‑session with /api/v2/memory/flush to persist the Markdown files.
7. Search – Query via /api/v2/memory/search. With only the LLM config you can use method: "keyword"; adding an embedding provider unlocks hybrid (vector + keyword) search.

Optional extensions

Feature	What you add	What you gain
Embeddings	[embedding] section in everos.toml	Hybrid search, memory reflection, skill extraction
Rerank	[rerank] config	Agentic search, Knowledge Wiki ranking
Multimodal ingestion	Install everos[multimodal]	Parse images, PDFs, audio, Office docs (requires LibreOffice for Office files)

Ecosystem & integrations

EverOS is positioned as the memory layer for a growing set of projects in the EverMind ecosystem:

• Reunite – semantic memory for family‑reconnection tools.
• Hive Orchestrator – browser‑native “hive‑mind” for multiple coding agents.
• AI Coding Assistants – long‑term memory for tools like evermemos-mcp.
• AI Data Technician, Study Buddy, Alzheimer’s Memory Assistant, and many more demos listed in the README showcase how the same memory backend can power personal assistants, data‑analysis agents, games, and wearable AI.

All of these are either linked repositories or external demos that consume EverOS’s HTTP API or Python client.

Development workflow

git clone https://github.com/EverMind-AI/EverOS.git
cd EverOS
uv sync               # set up a virtual env and install deps
uv run everos demo    # try the no‑key demo
uv run everos init    # add your OpenRouter key
make test             # run the test suite

The project uses uv for fast dependency management and provides a make test target for CI.

Who should use it?

• AI agent developers who need a portable, version‑controlled memory store without provisioning cloud databases.
• Researchers building long‑term personal assistants or multi‑agent systems and want their memory data to be inspectable and reproducible.
• Tool makers who want to add a “knowledge wiki” or reflective memory to existing LLM‑driven apps.

TL;DR

EverOS is a lightweight, local‑first memory runtime for LLM agents. It stores everything as editable Markdown, indexes it with SQLite + LanceDB, and exposes a simple REST API. With a single OpenRouter key you can start adding and recalling memories; optional embeddings, reranking, and multimodal parsers let you scale to richer retrieval and content types.