AgentOS — Local-first persistent memory for AI coding agents

Phase 0.5 MVP: a small Rust MCP server that gives Claude Code persistent, project-scoped memory using SQLite FTS5.

Why AgentOS?

AI coding agents lose important context when a session ends. Architecture decisions, failed approaches, project conventions, and bug-fix details often have to be explained again.

AgentOS provides a small local memory layer that coding agents can access through MCP. Memories are stored in SQLite and retrieved with FTS5 keyword search—without a cloud account, embedding API, or external database.

Phase 0.5 is intentionally narrow: save explicit project memories and retrieve them in future Claude Code sessions.

Features

• Persistent memory across Claude Code sessions
• Local SQLite storage with bundled SQLite
• Fast keyword retrieval through SQLite FTS5
• Two focused MCP tools: memory.remember and memory.search
• Project isolation using the MCP server's working directory
• MCP JSON-RPC communication over standard input and output
• No telemetry, network service, cloud database, or API key

Architecture

┌─────────────────────┐
│     Claude Code     │
└──────────┬──────────┘
           │
           │ MCP JSON-RPC
           │ newline-delimited stdio
           ▼
┌─────────────────────┐
│       AgentOS       │
│                     │
│  memory.remember    │
│  memory.search      │
└──────────┬──────────┘
           │
           │ rusqlite
           ▼
┌─────────────────────┐
│   SQLite + FTS5     │
│                     │
│  Local persistence  │
│  Keyword search     │
└─────────────────────┘

SQLite is the source of truth. The FTS5 index is kept synchronized with the memories table through SQLite triggers.

Quick Start

Requirements

AgentOS currently uses a Windows-first development workflow.

Install:

• Windows 10 or Windows 11
• Rust 1.80 or newer
• Visual Studio 2022 Build Tools
• The Desktop development with C++ workload
• MSVC v143 build tools
• Windows 10 or Windows 11 SDK
• Claude Code

A separate SQLite installation is not required. AgentOS compiles and links bundled SQLite through rusqlite.

1. Clone the repository

Open PowerShell:

git clone https://github.com/OfficialTanishSharma/agentos.git
Set-Location .\agentos

2. Verify the Rust toolchain

rustc --version
cargo --version
rustup show

The active host should normally be:

x86_64-pc-windows-msvc

If required, select it explicitly:

rustup default stable-x86_64-pc-windows-msvc

3. Test and build AgentOS

cargo test
cargo build --release

The release binary will be created at:

target\release\agentos.exe

Verify it:

Get-Item .\target\release\agentos.exe

4. Connect AgentOS to Claude Code

Resolve the release binary to an absolute path:

$agentos = (Resolve-Path .\target\release\agentos.exe).Path

Register it as a project-scoped stdio MCP server:

claude mcp add --transport stdio --scope project agentos -- $agentos

Inspect the configuration and connection status:

claude mcp get agentos
claude mcp list

The expected status is:

✔ Connected

If the server is waiting for project approval, start Claude Code and approve the MCP configuration:

claude

5. Save a memory

Inside Claude Code, ask:

Call memory.remember with these values:

title: AgentOS storage decision
body: AgentOS Phase 0.5 uses bundled SQLite with FTS5 for local persistent keyword search.
tags: architecture, sqlite, phase-0.5

6. Retrieve the memory

Ask:

Call memory.search with query "SQLite FTS5 storage" and limit 10.

Exit Claude Code, start a new session from the same project directory, and repeat the search. The saved memory should remain available.

Database location

On Windows, the default database is:

%USERPROFILE%\.agentos\agentos.db

Inspect it with PowerShell:

$db = "$env:USERPROFILE\.agentos\agentos.db"

Get-Item $db
Get-Item "$db-wal" -ErrorAction SilentlyContinue
Get-Item "$db-shm" -ErrorAction SilentlyContinue

Override the location for the current PowerShell session:

$env:AGENTOS_DB = "$PWD\agentos-test.db"

Remove the override:

Remove-Item Env:AGENTOS_DB

MCP Tools

memory.remember

Use memory.remember for information that should survive future coding sessions:

• Architecture decisions
• Bug fixes and root causes
• API contracts
• Project conventions
• Commands that solved a problem
• Failed approaches that should not be repeated

Example arguments:

{
  "title": "Use WAL mode for SQLite",
  "body": "AgentOS uses SQLite WAL mode so readers are not blocked by normal write activity. A five-second busy timeout handles short lock contention.",
  "tags": [
    "architecture",
    "sqlite",
    "concurrency"
  ]
}

Example result:

Remembered 'Use WAL mode for SQLite' with ID 46a71d9cb5b748efbd66738758cb089a.

Arguments:

memory.search

memory.search performs local FTS5 keyword search. Titles receive a higher BM25 ranking weight than memory bodies and tags.

Example arguments:

{
  "query": "SQLite WAL concurrency",
  "limit": 10
}

Example result:

1. Use WAL mode for SQLite
ID: 46a71d9cb5b748efbd66738758cb089a
Tags: architecture, sqlite, concurrency
Created: 2026-09-18T14:32:10.125Z
AgentOS uses SQLite WAL mode so readers are not blocked by normal write activity. A five-second busy timeout handles short lock contention.

Arguments:

Search terms are quoted and joined with OR. This keeps the FTS query safe and favors useful partial matches, but it is not semantic search.

How It Works

Claude Code starts AgentOS as a child process and communicates with it using newline-delimited JSON-RPC over stdin and stdout.

AgentOS implements the MCP methods required for this MVP:

initialize
ping
tools/list
tools/call

Protocol responses are written only to stdout. Startup information and diagnostic messages are written to stderr so they do not corrupt MCP framing.

When the process starts, AgentOS:

1. Resolves the database path.
2. Creates the database directory if it does not exist.
3. Opens SQLite with a five-second busy timeout.
4. Enables WAL journal mode.
5. Creates the memory table, FTS5 index, and synchronization triggers.
6. Resolves the current working directory as the default project key.
7. Waits for MCP requests on stdin.

The default project key is the canonical working-directory path with Windows path separators converted to forward slashes. Every search filters by this key.

Because project isolation depends on the working directory, Claude Code should be started from the same project root when memories are saved and retrieved.

What's NOT Included

Phase 0.5 does not include:

• Semantic search
• Embeddings or local language models
• LanceDB, Qdrant, or another vector database
• Automatic source-code indexing
• Git-history indexing
• Conversation import
• Cross-agent session handoffs
• Skill discovery or skill routing
• Background daemon or file watcher
• Memory editing or deletion MCP tools
• Memory deduplication
• Team synchronization
• Cloud backup
• HTTP transport
• TUI or desktop interface
• Telemetry

These limitations are intentional. Phase 0.5 tests the smallest useful version of persistent coding-agent memory before introducing additional storage and retrieval systems.

Development

Format

cargo fmt
cargo fmt --check

Static checks

cargo check
cargo clippy --all-targets --all-features -- -D warnings

Tests

cargo test
cargo test -- --nocapture

The current tests cover:

• Saving and retrieving a memory through FTS5
• Isolating search results by project key

Release build

cargo build --release

Generate a SHA-256 checksum:

Get-FileHash .\target\release\agentos.exe -Algorithm SHA256 |
    Format-List

Manual MCP smoke test

Create a JSONL request file:

@'
{"jsonrpc":"2.0","id":1,"method":"initialize","params":{"protocolVersion":"2024-11-05","capabilities":{},"clientInfo":{"name":"manual-test","version":"1.0"}}}
{"jsonrpc":"2.0","method":"notifications/initialized","params":{}}
{"jsonrpc":"2.0","id":2,"method":"tools/list","params":{}}
{"jsonrpc":"2.0","id":3,"method":"tools/call","params":{"name":"memory.remember","arguments":{"title":"Manual test","body":"AgentOS stored this memory through MCP JSON-RPC.","tags":["test","mcp"]}}}
{"jsonrpc":"2.0","id":4,"method":"tools/call","params":{"name":"memory.search","arguments":{"query":"manual MCP test","limit":10}}}
'@ | Set-Content .\requests.jsonl -Encoding utf8

Run it through AgentOS: