Open-source MCP server

Your agent forgets everything.
Now it won't.

memd gives AI coding agents long-term memory. Decisions, patterns, context — stored once, available in every future session.

Quick setup View source

session

~ memd_store key="auth-flow" type=decision

Stored → persists across all sessions.

~ memd_search query="auth" hybrid=true

3 results (38ms) — semantic + structured

~ memd_recall key="auth-flow"

JWT + refresh tokens, rate limit at gateway…

The problem

Every session starts from zero

Your AI agent makes great decisions — then forgets all of them. Next session, it re-discovers the same patterns, repeats the same mistakes, asks the same questions. You become the memory.

🔁

Repeated work

Same architecture discussions, every single time.

🧩

Lost context

Decisions made last week? Gone. Patterns discovered? Gone.

📋

Manual notes

You end up maintaining CLAUDE.md files by hand.

🐌

Slow ramp-up

Every conversation needs 5 minutes of "here's what we decided."

Capabilities

A complete memory system

Everything an agent needs to build, maintain, and query its own knowledge base.

Core

Structured context

Ten entry types — decisions, patterns, errors, schemas, constraints, solutions, and more. Each with priority, scope, tags, and TTL.

Hybrid retrieval

SQL filters and vector similarity search run in parallel, then merge results. Find knowledge by meaning, not just keywords.

State

Checkpoints

Snapshot development state — completed steps, blockers, next actions. Resume any task from exactly where it was left.

Audit

Progress logs

Append-only trail of every action. Files changed, outcomes, notes. Full session history that never gets lost.

Lifecycle

Sessions

Track working periods with start/end lifecycle. Context persists across sessions — the whole point.

Work

Persistent tasks

Track work items across sessions with status lifecycle, priority, and dependency blocking. Never lose your todo list between conversations.

Ops

Auto cleanup

TTL-based expiry sweeps stale entries. Priority-aware: critical knowledge is never auto-removed. Zero maintenance.

Under the hood

Three layers, clean separation

Layer 1

AI Agent

Claude Code · Copilot · Cursor

MCP

Layer 2

MCP Server

TypeScript · stdio transport

REST

Layer 3

memd API

Go · stateless

PostgreSQL

Structured storage & queries

Qdrant

Vector similarity search

Embedder

ONNX local · OpenAI fallback

Setup

Three steps. Two minutes.

Works with any MCP-compatible agent. Backend hosted at api.memd.dev.

STEP 01

Install and configure

$ npx @memd/mcp:latest --setup

# Registers your API key
# Backend: api.memd.dev

STEP 02

Wire up your agent

// .mcp.json
{
  "mcpServers": {
    "memd": {
      "command": "npx",
      "args": ["@memd/mcp"],
      "env": {
        "MEMD_API_URL": "https://api.memd.dev",
        "MEMD_API_KEY": "your-key"
      }
    }
  }
}

STEP 03

Start building

$ claude

# 19 memory tools ready.
# Your agent now remembers.

API surface

19 tools, one protocol

memd_store

Save a context entry with type, priority, tags, TTL

memd_recall

Retrieve a specific entry by its unique key

memd_search

Hybrid semantic + structured search

memd_list

Filter entries by type, scope, tags, priority

memd_update

Partial update to any existing entry

memd_delete

Remove an entry permanently

memd_checkpoint_*

Save, get, list, delete development snapshots

memd_progress_log

Append an action to the session audit trail

memd_session_*

Start and end working sessions

memd_task_create

Create a persistent task with priority and dependencies

memd_task_update

Update status, priority, or add/remove blockers

memd_task_list

List tasks filtered by status and priority

memd_task_get

Retrieve a specific task by ID

memd_task_delete

Remove a task permanently

Your agent forgets everything.Now it won't.