If you use more than one AI coding tool — and at this point, most developers do — you know the tax. Claude Code hits its context limit. You open Codex. You spend the next ten minutes re-explaining your project structure, your naming conventions, and the caching bug you fixed last Tuesday. The agent does not remember. It never did.
Beacon, an MIT-licensed, open source memory layer from Asymptote Labs, launched on October 20 with a $5.5M seed round and a direct answer to that problem. It captures sessions from 29+ coding harnesses — Claude Code, Cursor, Codex, OpenCode, Cline, GitHub Copilot — normalizes them into a local JSONL store, and makes the useful parts available to your next agent, whatever tool that turns out to be.
How It Works #
Beacon installs a lightweight capture hook alongside each supported harness. The hook records prompts, tool calls, file edits, approvals, and token usage in real time using OpenTelemetry-based normalization — same format regardless of which tool generated the session. That data lands in a local JSONL file you own.
The important part: Beacon does not automatically inject anything back into your agents. This is a deliberate design choice and the right one. You browse captured sessions with beacon traces, optionally run a Jev or SubQ evaluator to score candidates, then write and approve each lesson manually. Once approved, lessons are served to any MCP-compatible agent via beacon mcp connect, or packaged as Agent Skills.
Nothing reaches your agents until you say it does. That makes Beacon slower than fully automated memory tools — and more reliable for professional codebases.
Beacon vs. claude-mem: Pick the Right Tool #
ByteIota covered claude-mem two days ago — the 108K-star Apache 2.0 memory plugin that hooks into Claude Code and auto-compresses sessions for semantic retrieval. If your entire workflow lives inside Claude Code, claude-mem is simpler and probably sufficient. It requires no manual lesson approval and works out of the box.
Beacon is for a different situation: you use multiple agents interchangeably. Insights from your Cursor sessions should improve your Claude Code sessions, and vice versa. No single-harness memory tool can do that. If you are context-switching between two or three coding agents on a shared repository — which is increasingly common — Beacon is the only free option that bridges them.
Getting Started #
Install on macOS:
brew update && brew trust asymptote-labs/tap && brew tap asymptote-labs/tap && brew install beacon
Linux gets an automated installer with SHA256 verification. Windows gets an x64 MSI with silent install support.
After install, run beacon mcp connect to register the MCP server with your agents. Point any MCP-compatible agent to Beacon’s endpoint and it will retrieve approved lessons at the start of each session.
One warning before you start: the default setup selects Cloud storage and backfills the last 30 days of captured sessions. If you are on an NDA-protected repository or have any concern about session data leaving your machine, choose Local storage and pass --no-backfill during setup. The option is there; it is just not the default.
The $5.5M Signal #
The seed round was led by Tower Research Ventures, with Lytical Ventures, 1984 Ventures, and South Park Commons participating. Tower’s own security team is running Beacon in production — that is a more meaningful endorsement than a press release. The GitHub repo hit 460% star growth in September and Beacon took number one Product of the Day on Product Hunt at launch.
CEO Justin D’Souza framed the broader bet: “AI agents are becoming a new operating layer inside every company, but infrastructure to understand and control them hasn’t caught up.” The MIT license keeps the base layer free. A Pro tier at $20 per month adds unlimited findings for teams that need it.
Who Should Adopt Now #
If you actively switch between two or more coding agents on the same repositories, Beacon solves a real and present problem. Install it, run a few sessions, use beacon traces to find one or two high-signal corrections worth preserving, write the lessons, and connect via MCP. The setup takes under ten minutes.
If you work exclusively in Claude Code and have no plans to change, the alternatives review at Era of Minds suggests claude-mem is still the stronger pick for single-harness workflows. Beacon would add overhead without benefit.
Agent memory is becoming infrastructure. The harness-agnostic layer is the harder and more valuable piece to build — and it is the one that was missing until now.