What Is beacon.sh? The Beginner's Guide to Shared Memory for AI Coding Agents

CodexFuse Team · 2026-11-04T19:24:03.228Z · 14 min read min read

What Is beacon.sh? The Beginner's Guide to Shared Memory for AI Coding Agents

If you have been wondering what is beacon.sh, here is the short answer: beacon.sh is an open-source, self-improving memory layer for AI coding agents. It captures your agent sessions across Claude Code, Cursor, Codex, OpenCode, and more than 20 other harnesses, then turns the useful workflows, corrections, and debugging patterns into reusable knowledge that future agents can pull in — no matter which harness you use next. The company behind it, Asymptote Labs, describes Beacon with one line: "A problem solved by one agent shouldn't need to be learned from scratch by another." (github.com/asymptote-labs/agent-beacon) This beacon.sh explained guide covers what the tool captures, how its review loop works, how to install it locally or with Beacon Cloud, what it costs, and where its coverage still has gaps.

If you are setting up a Linux box to experiment on, start with our VPS hosting guide — a small always-on server is the natural place to run agent harnesses overnight while Beacon quietly records what they learn.

What Is beacon.sh, Exactly?

Beacon is a cross-harness telemetry and memory layer for coding agents. "Cross-harness" is the key phrase: instead of each AI tool keeping its own isolated session history, Beacon sits underneath all of them and normalizes everything into one shared event model, so knowledge learned in one tool is visible to the others. The official site puts it this way: "Beacon turns agent sessions across Claude Code, Cursor, Codex, OpenCode, and 20+ other harnesses into a compounding knowledge layer, so every agent starts smarter than the last." (beacon.sh) A knowledge layer "compounds" when every new session starts with everything the past sessions learned — the opposite of the daily amnesia every agent user knows.

One disambiguation before anything else, because it will save you confusion: beacon.sh is not the only product named Beacon. Canyon Road's "Beacon" is supervised endpoint AI security — "Beacon provides visibility and guardrails for AI tools running on employee endpoints (macOS and Windows), with policies enforced at the moment actions execute" — a different product from Asymptote Labs' beacon.sh open-source agent memory layer. (canyonroad.ai, crawled 2026-11-04) This guide is about the Asymptote Labs one.

Why Do AI Coding Agents Need Shared Memory?

They need it because every harness is an island. You fix a subtle bug in Claude Code on Monday; on Tuesday you open Codex or Cursor and correct the exact same mistake again, because each harness keeps its own memory. The dev.to guide on Beacon opens with exactly that scene: "You fix a mistake in Claude Code. Tomorrow you open Codex or Cursor and correct the exact same mistake again, because each harness keeps its own memory." (dev.to/dedupedai, 2026-10-25) Beacon's founders say they built the tool out of that exact frustration.

The official site frames the problem bluntly: "your coding agents forget everything the second the session ends. Every fix. Every dead end you already hit. Every 'no, not like that.' Gone. Tomorrow you teach it all over again." (beacon.sh) The fix Beacon proposes is structural, not magical: beacon cross-harness memory means knowledge learned in Cursor can inform Codex, and a Codex session can inform OpenCode. Knowledge moves; the harnesses stay replaceable.

How Does Beacon Turn a Session Into Memory? The Dreaming Workflow

Beacon converts raw sessions into reusable memory through a loop the project calls "dreaming": capture a finished session, evaluate what in it worked, extract the useful pattern, have a human review and approve it, then reuse it across future agents. The README states the loop verbatim as "Run agents ↓ Capture session history ↓ Evaluate what worked ↓ Extract useful knowledge ↓ Review + approve ↓ Reuse across future agents." (github.com/asymptote-labs/agent-beacon) The review step is the design's conscience — nothing becomes shared knowledge until a person signs off on it, which is how Beacon keeps a hallucinated workaround from becoming permanent lore.

Here is the concrete arc the company's own documentation uses, retold from the Automater piece on the seed round:

1. An agent fixes a webhook retry bug. During the session it discovers a repository-specific signing rule the hard way.
2. Beacon reviews the finished session. Its dreaming step picks out the correction — the signing rule — as the thing that actually mattered.
3. A draft skill is created. The fix is distilled into a concrete, reusable skill rather than left buried in a transcript.
4. A human reviews it once. You approve (or reject) the draft before it ships into shared memory.
5. Any later agent can retrieve it. The next time the same question comes up, a Codex or Cursor session pulls the skill in through MCP or Agent Skills — the fix is explained once, by the past, instead of rediscovered from scratch. (automater.ai, 2026-10-21)

Say this once, because it is the single most important mental model: Beacon's memory is distilled and reviewed knowledge, not a transcript dump. The value is not that it remembers everything — it is that a human curates what matters and every future agent inherits it. Expect to curate what Beacon extracts rather than get perfect memory on day one.

What Does Beacon Actually Capture?

Beacon captures agent execution where it happens and normalizes it into a common OpenTelemetry-based event model — the beacon opentelemetry agents pipeline that makes one capture format legible everywhere. The README is explicit about the fields: sessions, prompts and responses, tool calls, commands, file activity, approvals, MCP interactions, and token usage. (github.com/asymptote-labs/agent-beacon) Using OpenTelemetry as the format — instead of inventing a proprietary schema — is why Beacon can also forward the same history to tools you already run, like Splunk, Datadog, or Elastic.

The README's coverage table lists roughly 30 local agent harnesses (Claude Code, Codex CLI, Cursor, Cline, OpenCode, Gemini CLI, GitHub Copilot CLI, and others) plus browser chats (Claude.ai, ChatGPT), cloud agents (Claude Code Cloud, Cursor Cloud, Devin Cloud), and SDK instrumentation (Anthropic, OpenAI, Vercel AI SDK). But coverage is not uniform, and the gaps are worth knowing before you install. This table summarizes the verified gaps from the README's coverage matrix:

The table above is compiled from the README's own coverage matrix as fetched on 2026-11-04, cross-checked against the Pivot News piece. (github.com/asymptote-labs/agent-beacon; pivotnews.ai, 2026-10-23) The honest takeaway: Claude Code gets the richest capture; every other harness is a partial view. Do not assume a memory built from Cursor sessions knows the token cost of what it did.

How Do You Install Beacon.sh?

This beacon.sh install walkthrough covers all three OS routes, and it takes minutes. On macOS you install through the Asymptote Labs Homebrew tap:

brew update
brew trust asymptote-labs/tap
brew tap asymptote-labs/tap
brew install beacon
beacon endpoint install

The Linux route is a curl-to-bash installer from the project's releases page (it verifies against the release's checksums.txt), and Windows uses an MSI:

curl -fsSL https://github.com/asymptote-labs/agent-beacon/releases/latest/download/install.sh | bash

msiexec /i BeaconEndpointAgent-<version>-x64.msi

(Install commands verbatim from dev.to/dedupedai, 2026-10-25, and beacon.sh.) The fastest path of all is to hand the job to an agent itself: the official site gives you one prompt — Read https://beacon.sh/start.md then set me up on Beacon — and lets your coding agent run the wizard. (beacon.sh)

If you install from source instead, you will want our Git beginner guide handy; if you would rather test in a throwaway container first, our Docker guide is the faster route.

Local or Beacon Cloud: Which Should You Pick?

Pick Local if you want everything on your machine, and Beacon Cloud if you want hosted memory with retrieval from anywhere. The choice matters because it decides where your agent traces live — and Beacon is unambiguous about the default: "Do not enable a destination the user did not choose. Beacon records locally by default and uploading is opt-in." (beacon.sh/start.md)

Local mode writes endpoint telemetry to durable JSONL on your own disk — inspectable with beacon traces in the terminal, a local dashboard via beacon endpoint dashboard, or raw at ~/.beacon/endpoint/logs/runtime.jsonl, which is plain JSONL "so you are never dependent on the UI." (dev.to/dedupedai, 2026-10-25) The local MCP server, beacon mcp serve, "never touches the network." Cloud mode, by contrast, is chosen during the interactive setup: signing in through beacon.sh preselects Beacon Cloud with an explicit Local opt-out, and confirming Cloud uploads the last 30 days of agent sessions stored on your machine. (beacon.sh/start.md; dev.to) One asymmetry to remember: local memory is project-scoped — lessons stay inside the repository where they were learned and are not automatically shared across unrelated projects. (aipricing.guru, 2026-10-22)

Wherever traces live, the installer also drops Beacon's Agent Skills into ~/.agents/skills and, when Claude Code is present, ~/.claude/skills — and only skills Beacon itself installed are ever replaced, "so a skill of the same name the user wrote is left alone." (beacon.sh/start.md) Disconnect any time with beacon endpoint disconnect.

How Much Does Beacon Cost?

Beacon's local path is free and open source; hosted and enterprise features cost money. The three setups on the official start page are: Beacon Cloud (recommended), Beacon Cloud Pro at $20/month (or $15/month billed annually — "Beacon Cloud plus unlimited findings, full MCP access ... memories agents can recall, and the full evidence trail on every finding"), and Local ("sessions stay on this machine and the user browses them in a local dashboard. Nothing is uploaded."). (beacon.sh/start.md) The broader tier map, per the pricing analysis, distinguishes local Open Source, Beacon Cloud, paid Beacon Cloud Pro, and Enterprise. (aipricing.guru, 2026-10-22)

Two money facts worth internalizing. First, the split: "Beacon itself is free telemetry. Asymptote's revenue comes from an enterprise platform built on top of it that intercepts live agent activity and enforces policy-driven guardrails at the prompt and tool level." (automater.ai, 2026-10-21) Second, the boundary: "Hosted memory over MCP is a Pro/Enterprise feature; the local memory path is separate." (aipricing.guru, 2026-10-22) And one honest non-claim: the same pricing review notes that "Neither the HN post nor the reviewed documentation establishes a measured percentage reduction in token spending" — Beacon does not promise cheaper inference or a way around model rate limits, and you should distrust anyone who says it does.

Beacon vs Claude-Mem vs Toli: Which Memory Tool Fits?

The memory-tool landscape splits by scope, and the three names keep different jobs. Claude-Mem is a single-harness memory tool — memory inside one agent's world. Beacon is the cross-harness layer: its whole point is that memory survives the switch between Claude Code, Codex, Cursor, and the rest. Toli is the A2A-protocol lesson — a different lane about agents talking to agents, not about what they remember. Drawing the line once: none of them rivals Beacon, because Beacon is the only one whose job is shared memory across harnesses.

If you run one agent and never switch, Claude-Mem's simplicity may be all you need. The moment you bounce between two or more harnesses — because of usage limits, a better model dropping, or just preference — Beacon is the tool built for that exact wall. (Drawn from the queue's sibling check: 1367 "How to Use Toli" — the A2A-protocol lesson; 1111 "How to Use Claude-Mem" — single-harness memory; 1363 "How to Use Graphiti" — temporal knowledge-graph memory. None is live on CodexFuse at this writing.)

Who Built Beacon, and Why Does It Matter Now?

Beacon is built by Asymptote Labs, a New York startup co-founded by former Doppel engineers Justin D'Souza and Shukan Shah. On October 20, 2026, the company announced an oversubscribed $5.5 million seed round led by Tower Research Ventures — the venture arm of Tower Research Capital — with Lytical Ventures, 1984 Ventures, and South Park Commons also participating. The beacon asymptote labs seed announcement was covered in detail by Automater the next day. (automater.ai, 2026-10-21) The repo, asymptote-labs/agent-beacon, is MIT-licensed, was created May 12, 2026, and carried 3,274 stars when checked on November 4, 2026.

The launch sequence explains the timing: the founders introduced the project on Hacker News on October 21, 2026 — "Earlier this year, my co-founder and I found ourselves frequently switching between different coding agents ... We became increasingly frustrated with having to re-explain our tasks, workflows, common mistake patterns, etc. anytime we switched agents. We found no good free alternatives out there, so we built our own and fully open-sourced it!" (Ask HN mirror, 2026-10-21) — and Beacon launched on Product Hunt in early November 2026, where a third-party listing recorded a score of 8.56, 444 votes, and a #4 hot-list placement. (aihot0818.com, crawled 2026-11-04) The founders' own framing is deliberately grand: "We're betting our $5.5M seed round that the next leap in AI coding won't come from OpenAI or Anthropic. It'll come from a new feedback loop that learns from you." (beacon.sh)

What Should You Watch Out For?

Four honest caveats before you install. First, the coverage gaps are real: memory built from Codex CLI sessions has no file-activity record, and Cursor sessions have no token record — the memory is only as complete as the capture. Second, dashboard lenses run in a sandboxed frame, but the docs warn that "the sandbox cannot block every way a browser can send data out, so only lenses the user trusts with their traces should be installed." (pivotnews.ai, 2026-10-23) Third, your traces contain whatever you pasted into an agent — API keys included — so Local mode is the right default if that bothers you, and reviewing what Beacon captured (beacon traces) is worth doing early. Fourth, Beacon is not a cheaper-model trick: nothing in the docs or the HN post establishes token savings, and Beacon does not transfer subscription allowances or reset rate limits. (aipricing.guru, 2026-10-22)

Sources

• beacon.sh — official site (tagline, problem framing, install prompt, seed-round launch quote). Crawled 2026-11-04.
• beacon.sh/start.md — official setup guide (plans, Local vs Cloud, skills paths, opt-in upload, MCP). Crawled 2026-11-04.
• github.com/asymptote-labs/agent-beacon — README (summary, MIT license, coverage matrix, dreaming workflow, OpenTelemetry event model, local JSONL). Crawled 2026-11-04.
• automater.ai — "Asymptote Labs Raises $5.5M Seed for Beacon, a Cross-Harness Memory Layer for Coding Agents" (published 2026-10-21; seed details, founders, dreaming workflow, webhook example, free-vs-paid split). Crawled 2026-11-04.
• pivotnews.ai — "Beacon ships cross-tool agent memory on a reported $5.5M" (published 2026-10-23; installer behavior, opt-in forwarding, lens sandbox caveat). Crawled 2026-11-04.
• dev.to/dedupedai — "How to share memory between Claude Code, Codex and Cursor with Beacon" (published 2026-10-25; install commands, Local vs Cloud, JSONL traces, SIEM forwarding). Crawled 2026-11-04.
• aipricing.guru — "Beacon Agent Memory: Costs, Privacy & Cross-Tool Use" (sources checked 2026-10-22; plan tiers, hosted-MCP boundary, project-scoped memory, no savings claim). Crawled 2026-11-04.
• aihot0818.com — beacon.sh tool listing (Product Hunt score, votes, first-listed date). Crawled 2026-11-04.
• producthunt.com/products/asymptotelabs — Beacon launch page (community blurbs, daily top-post badge). Crawled 2026-11-04.
• Ask HN mirror (ask.rivestack.io) — founders' October 21, 2026 launch post (frustration, open-source motivation). Crawled 2026-11-04.
• canyonroad.ai — "Breaking the Agentic Kill Chain" (Beacon disambiguation line). Crawled 2026-11-04.

Frequently Asked Questions

Is beacon.sh free?

Yes — the local path is free and the project is MIT-licensed open source; the official site says "Free to get started · No credit card required." (beacon.sh) Paid tiers start with Beacon Cloud Pro at $20/month or $15/month billed annually, and hosted memory over MCP is a Pro/Enterprise feature. (beacon.sh/start.md; aipricing.guru)

Which AI coding agents does Beacon work with?

The README's coverage table lists roughly 30 local harnesses — Claude Code, Codex CLI, Cursor, Cline, OpenCode, Gemini CLI, GitHub Copilot CLI, and others — plus browser chats (Claude.ai, ChatGPT), cloud agents (Claude Code Cloud, Cursor Cloud, Devin Cloud), and SDK instrumentation for Anthropic, OpenAI, and the Vercel AI SDK. (github.com/asymptote-labs/agent-beacon) Claude Code gets full capture; other harnesses have documented gaps.

Does Beacon upload my code to the cloud?

Not unless you choose it. Beacon records locally by default, writes traces to JSONL on your machine, and uploading is opt-in — Local mode keeps everything on the device. (beacon.sh/start.md) Interactive setup preselects Beacon Cloud with an explicit Local opt-out, and confirming Cloud uploads the last 30 days of sessions. (dev.to/dedupedai)

What is the difference between beacon.sh and Canyon Road Beacon?

They are different products that share a name. Asymptote Labs' beacon.sh is an open-source, local-first memory layer for AI coding agents. Canyon Road's "Beacon" is supervised endpoint AI security with policies enforced on employee endpoints. (canyonroad.ai, 2026-11-04)

Can Beacon replace Claude-Mem or Toli?

It does not need to. Claude-Mem is single-harness memory, Toli is about the A2A protocol for agent-to-agent communication, and Beacon is the cross-harness shared-memory layer — different jobs, no rivalry. If you regularly switch between Claude Code, Codex, and Cursor, Beacon is the tool built for that wall.