Your agents keep forgetting; Beacon keeps the receipts

What it does

Beacon is a Go-based memory layer for AI coding agents. It captures full session history — prompts, responses, tool calls, commands, file edits, approvals, MCP activity, token usage — across Claude Code, Cursor, Codex, OpenCode, Cline, and a long tail of other harnesses, then normalizes everything into one OpenTelemetry-based event model. The pitch: a problem solved in Cursor shouldn’t have to be re-learned by Codex.

The interesting bit

It sits across the harness layer rather than inside any one of them, so knowledge compounds as you switch tools. Extracted workflows, corrections, and debugging patterns go through a review-and-approve step before becoming shared memory that future agents retrieve via MCP or Agent Skills — a human gate on what counts as “learned.”

Key highlights

• One trace format for 25+ local agents, plus browser chats (Claude.ai, ChatGPT), cloud agents, CI jobs, and SDK instrumentation
• Local-first: durable JSONL on disk, a read-only local dashboard, and no dependency on the UI
• Forwarding to Splunk, Datadog, Elastic, Sentinel, Sumo, Wazuh, S3, and other infra you already run
• Exact session replay — approvals and MCP interactions included, not just prompts
• Explicit privacy modes; signing in alone forwards nothing

Caveats

• Coverage is uneven: several harnesses lack approvals, MCP, or token capture (see the README’s per-runtime table), and VS Code entries are partial
• Interactive setup preselects the hosted “Beacon Managed” option — you have to actively choose Local, though the confirm screen discloses this
• The knowledge-extraction loop depends on you reviewing and approving; it’s not automatic compounding

Verdict

Worth a look if you bounce between multiple agent harnesses and hate re-teaching each one your repo’s conventions. If you live in a single tool, your harness’s built-in memory probably covers most of this already.