Persistent Memory for Claude Code

Claude Code forgets everything between sessions except what lives in CLAUDE.md and your repo files. If you want Claude Code to remember conventions, decisions, and fixes across sessions — and across projects — you need a real memory layer.

AI DevKit provides one: a local SQLite memory service that Claude Code can read and write through MCP. Nothing leaves your machine.

CLAUDE.md is loaded into every session, which makes it great for stable project rules — and a poor fit for everything else:

• It grows until it crowds out real context in the prompt window.
• It mixes project rules with personal knowledge that does not belong in git.
• It cannot be searched; the agent sees all of it or none of it.

A memory store inverts that: knowledge is saved once and retrieved only when relevant to the current task.

Run machine setup once. It wires the ai-devkit-memory MCP server into Claude Code's global MCP config automatically:

npm install -g ai-devkit
ai-devkit setup

Or without a global install:

npx ai-devkit@latest setup

Confirm the wiring with ai-devkit status, then restart Claude Code.

Once the MCP server is connected, just ask:

Storing:

"Remember that all API responses must handle BigInt serialization. Save it to memory with the tag #backend."

Retrieving:

"Check memory for any API standards I need to follow before building this endpoint."

Claude Code calls the memory tools on its own, so knowledge from last week — or from a different project in the same scope — surfaces when it matters.

You can also manage memory directly:

# Store a decision
ai-devkit memory store \
  --title "Use BigInt for API responses" \
  --content "Serialize BigInt with BigInt.toString() before sending API responses containing large numbers." \
  --tags "backend,api" \
  --scope "repo:myorg/my-repo"

# Search it later
ai-devkit memory search --query "api serialization"

Search is hybrid — lexical matching plus local semantic embeddings — so related entries surface even when the wording differs. See Memory for the full command reference.

Scope	Applies to
global (default)	Every project on this machine
project:<name>	One named project
repo:<org/repo>	One git repository

Agents automatically prioritize knowledge matching the scope of your current working directory.

Memory is stored in a local SQLite file at ~/.ai-devkit/memory.db (configurable via memory.path in .ai-devkit.json). It is never sent to a cloud service, and because it is a single file you can copy it to another machine.

The same memory server is wired into every MCP-capable agent AI DevKit detects — Cursor, Codex, Gemini CLI, opencode, and more. A convention Claude Code saves today is available to Codex tomorrow.

• Memory documentation
• The memory skill
• Supported AI agents and environments
• Why AI coding agents forget context