EverOS: Open-Source Memory Layer for AI Agents Using Markdown, SQLite, and LanceDB

AI agents often lose context when a new session starts, forcing users to repeat project background, debugging history, and preferences. EverOS, an open-source memory plugin for DeepSeek Harness (also integrated with Hermes and OpenClaw), addresses this by providing a persistent, cross-session memory layer that has garnered 13.5k stars on GitHub.

Architecture: Markdown + SQLite + LanceDB

EverOS separates memory storage, state management, and retrieval into three distinct components:

Markdown files store raw memory content. Users can open, inspect, and manually edit these files, and they can be version-controlled with Git, avoiding lock-in to a proprietary UI.

SQLite manages system state, synchronization queues, and audit logs, ensuring orderly memory writes and index updates.

LanceDB builds retrieval indexes supporting vector search, full-text search, and conditional filtering. This allows semantic matching even when queries use different wording from stored memories.

This division lets users directly control memory files — if an AI records an incorrect project convention, the user can correct the Markdown file and the system will sync the updated indexes.

Two Types of Memory: User-Side vs. Agent Experience

EverOS distinguishes between:

User memory — answers "Who is this user? How do they want tasks done?" (e.g., project conventions, interface standards).

Agent experience memory — answers "What has the agent done before? Which methods worked? What can be reused?" (e.g., a discovered dependency compatibility issue during debugging).

Managing these separately lets the system inject both project requirements and past execution lessons into future tasks.

Reflection and Memory Evolution

Storing every conversation verbatim leads to duplication, outdated conclusions, and contradictory records that waste context and confuse the model. EverOS introduces a background reflection and evolution mechanism that periodically consolidates memories: merging related experiences, updating user profiles, and extracting reusable skills.

Integration with DeepSeek Harness

Installation steps (requires Python 3.12+):

# 1. Install and initialize EverOS
uv tool install everos
everos init

# 2. Configure LLM key (for extraction/advanced retrieval; OpenRouter-compatible endpoints work)
# Edit ~/.everos/everos.toml or set environment variables

# 3. Start EverOS service (default local port 8000)
everos server start

# 4. Install the DSH plugin
dsh plugin --profile web add @everos-ai/dsh-plugin

EverOS aims to turn accumulated information into callable resources for the next task, making long-running development projects, personal knowledge bases, and research workflows more efficient by reducing dependence on a single application and enabling flexible memory migration and maintenance.