Beads

Overview

Beads addresses the same problem as claude-mem — coding agents forget everything between sessions — but takes a fundamentally different design path. Where claude-mem is a Claude Code plugin that ships compression via the Anthropic Agent SDK and ties tightly to one runtime, Beads is a standalone Go binary that any compliant agent can talk to via its MCP server interface. That decision unlocks cross-agent memory: a session in Claude Code on Monday and a session in Codex on Tuesday can both read and write the same memory chain.

The repo crossed 22K stars on April 27, 2026 and was the third-highest-velocity agent-tooling project of the day at +485 stars (behind warpdotdev/warp and mattpocock/skills). The Go-binary distribution model is unusual in this category — most agent tooling ships as TypeScript or Python packages — but it’s a deliberate choice. Beads can run on a developer’s laptop, a CI runner, a remote dev server, or a teammate’s machine without runtime drift.

Why It Matters

The “memory layer” for coding agents is rapidly bifurcating. One path is runtime-bound memory — claude-mem for Claude Code, similar plugins emerging for Codex and Cursor — where each agent runtime gets its own memory implementation. The other path is runtime-agnostic memory, exposed via standard interfaces (MCP, OpenAPI). Beads is the strongest open-source entry on the runtime-agnostic path, which makes it durable in a market where developers run multiple agents and can’t afford to maintain N separate memory stores.

Beads also picks up a thread that the broader skills/harness trend has been pulling at: the value moves from the model to the layer around the model. If the harness, skills, and memory are portable across agents, vendor lock-in collapses. Beads is one of the primitives that makes that portability real.

Key Capabilities

Single-binary distribution: brew install beads or download from releases. No Python, no Node, no Docker. Starts in milliseconds, sub-50MB memory footprint, runs everywhere Go does.

Cross-agent compatibility: Beads exposes its memory store via an MCP (Model Context Protocol) server. Any MCP-compliant agent — Claude Code, Codex, OpenCode, Cursor with the right configuration, Aider via MCP bridge — can read and write the same beads.

Session compression: At the end of each session, Beads compresses the conversation transcript, file diffs, and tool calls into a ‘bead’ — a structured memory object indexed by relevance, time, topic, and file paths touched. The compression strategy is configurable; the default uses LLM summarization with a local fallback.

Project-scoped chains: Memory is organized per-repository. Beads in your ~/work/api project never leak into your ~/work/frontend project. Within a repo, beads can be filtered by branch, by date range, or by topic tag.

Local-first storage: All memory stored in .beads/ at the project root or in ~/.beads/ for global memory. SQLite-backed, no external service, no telemetry.

Getting Started

# Install
brew install gastownhall/tap/beads
# Initialize in a project
cd ~/work/api && beads init
# Start the MCP server
beads serve

Configure your agent (Claude Code, Codex, etc.) to connect to the local MCP server at localhost:7777. From there, beads accumulate automatically as you work.

Who It’s For

Developers running multiple AI coding agents who don’t want to maintain N separate memory stores. Teams that share a project memory across coworkers (Beads chains can be committed to git or synced via cloud storage). Anyone burned by losing context in a long debugging session who wants explicit, queryable, version-controlled memory rather than a black-box plugin.