Beacon: Open-Source Memory Layer for AI Coding Agents

Beacon is an open-source, cross-harness memory layer for AI coding agents, built after its authors got tired of re-explaining the same tasks when switching between Claude Code, Codex, and Cursor. It hooks into 20+ agent harnesses, records session telemetry, and turns it into reusable skills exposed over MCP.

What it actually captures

• Session telemetry in one normalized format: prompts, tool calls, command executions, and other events from any supported agent harness.
• Distillation of that raw telemetry into common patterns and repeatable workflows.
• A human-approval step before anything is persisted, so the distilled knowledge gets reviewed rather than blindly written back.
• Exposure to future agents through skills and MCP, so the next harness you switch to starts with the same context.

Why a separate layer

Memory built into a single harness only travels as far as that harness. If you move between Claude Code, Codex, and Cursor — which the authors say they did constantly when hitting shaky usage limits or a newer model dropped — each one starts from zero. Beacon sits outside any single vendor's ecosystem and keeps the session history in a format your agents can consume regardless of which frontend is driving the model.

The stated design position: agent data is yours to own, and shouldn't be locked behind a frontier lab's closed memory infrastructure. The project is fully open source.

Setup

The source doesn't include install commands or a config sample — check the repo for the current harness list and setup steps:

git clone https://github.com/Asymptote-Labs/agent-beacon

From there, the practical loop is: point a supported harness at Beacon, work as usual, then review the distilled patterns it proposes and approve the ones worth keeping as skills.

Things to evaluate before adopting

• Telemetry capture means prompts and shell commands land on disk — confirm retention and where that store lives.
• Distillation quality matters more than capture volume; the approval step exists precisely because auto-extracted "workflows" can be noise.
• MCP as the delivery mechanism is a reasonable bet given how many harnesses already speak it, but the per-harness hook coverage is what you'll actually feel day to day.

It's aimed at developers running more than one coding agent, or one agent across more than one project, who are tired of restating conventions every session.