Beacon Review: Free Memory for Every AI Coding Agent

Your coding agent forgets everything the moment a session ends. The fix you walked it through yesterday, the project rule you explained twice, the dead end you already hit — all gone. Tomorrow you teach it all over again, and you pay for the tokens to do it.

Beacon, from New York startup Asymptote Labs, is an open-source memory layer built for exactly that problem. It records your agent sessions across Claude Code, Cursor, Codex, OpenCode and 20+ other coding tools, then turns the useful parts into reusable knowledge that future agents can pull back. The project launched in May 2026, raised a $5.5 million seed round on October 20, and hit the top of Product Hunt on November 2. As of November 7 it holds 3,363 stars on GitHub.

This review covers what Beacon records, how setup works, what the free tier actually buys, and the catches that matter before you install it.

What Beacon actually does

Beacon sits across your coding agents instead of inside any one of them. Every session it captures gets normalized into one OpenTelemetry-based event model: prompts, responses, tool calls, commands, file activity, approvals, MCP interactions and token usage. Instead of a separate proprietary history per tool, you get a single dataset you can inspect, search and build on.

The point is portability. A debugging fix you work out in Claude Code becomes a skill your Cursor session can retrieve the next day, and a correction you give Codex does not have to be repeated in OpenCode. The README states the idea in one line: a problem solved by one agent should not need to be learned from scratch by another.

Beacon turns that history into memory through a loop it documents openly: run your agents, capture the session history, evaluate what worked, extract the useful knowledge, review and approve it, then reuse it across future agents. Beacon says it uses Jev, OpenAI's decision model, to score which sessions are worth learning from, so the review step surfaces corrections and debugging patterns rather than every keystroke. Agents retrieve the result through MCP or Agent Skills.

You can browse what was captured locally with beacon traces, open the local dashboard with beacon endpoint dashboard, or read the raw event stream as JSONL at ~/.beacon/endpoint/logs/runtime.jsonl. By default nothing leaves your machine: local JSONL is the default destination, and you can also forward the same normalized history into Splunk, Datadog, Elastic, Microsoft Sentinel, CrowdStrike Falcon LogScale, Sumo Logic, Wazuh, AWS S3, Google Cloud Storage or CloudWatch.

Which agents it covers, and where coverage thins

The README's support matrix lists 29 local runtimes, including Claude Code, Claude Cowork, Cursor, Codex CLI and Desktop, OpenCode, Cline, Devin, Gemini CLI, GitHub Copilot CLI, Kiro, Kimi Code, Qwen Code, OpenHands, Antigravity, Factory Droid, Vercel's fx, goose, Grok Build, Hermes Agent, Muse Code, Oh My Pi, OpenClaw Gateway, Pi, Prime Agent, Senpi and VS Code. Browser chat on Claude.ai and ChatGPT is covered through an extension, and cloud agents from Claude Code, Cursor and Devin are covered too, along with CI jobs.

Coverage is not uniform, and this is the first thing to check before installing. The matrix shows, for example, that Cursor records no token usage, Codex CLI and Desktop record no file activity, Gemini CLI records no session at all, and Devin Desktop records no session either. OpenClaw's token capture is partial. If your workflow depends on one of those gaps, the memory Beacon builds for that harness will have holes in it.

Setup: Local or Cloud, and the choice that matters

Setup takes minutes. Paste one line into your coding agent: Read https://beacon.sh/start.md then set me up on Beacon. The manual route is short too. On macOS it is brew tap asymptote-labs/tap, brew install beacon, then beacon endpoint install. Linux uses an install script that downloads the package for your architecture and verifies it against the release checksums file; the README tells you to read the script first or verify the build-provenance attestation instead of piping it straight into your shell. Windows uses an MSI.

The setup decision that matters is the destination, and the installer does not hide it. Interactive setup signs in through beacon.sh and preselects Beacon Cloud. Signing in forwards nothing; confirming the Cloud option starts forwarding your history to Asymptote Labs, and the confirm screen says so before you accept. Local is the explicit opt-out, and you can disconnect at any time with beacon endpoint disconnect.

Pick Local if you want everything to stay on your machine. Pick Cloud if you want the hosted dashboard, session search over MCP and link sharing. The common mistake is clicking through the preselected Cloud option without reading the confirm screen, then being surprised that session history is leaving the machine. Beacon records prompts, responses, commands and file edits, so an API key pasted into a prompt can end up in a trace either way.

Setup also installs the Beacon Skills into ~/.agents/skills and ~/.claude/skills unless you turn that off, and on Cloud it registers the beacon-managed MCP server in Claude Code, Codex CLI, Cursor, VS Code, Gemini CLI and OpenCode with beacon mcp connect. Each harness signs in with OAuth the first time, no token is written, and beacon mcp disconnect removes exactly what was added. The local beacon MCP server is separate and never touches the network.

Pricing: what free buys, and where Pro starts

Beacon Cloud is free to start with no credit card required, and the free tier is usable rather than a demo. It covers 20+ agents, every agent and machine in one timeline, token-spend visibility, 50 findings, one-click skills and fix prompts, session search over MCP, and sharing any session with a link.

At November 7, 2026 rates, Pro at $20 a month works out to about £14.94 or C$28.23. Team at $30 a seat is about £22.41 or C$42.35 per seat, plus the $10 base fee at about £7.47 or C$14.12. Those are approximations from 1 USD = 0.7471 GBP and 1.4117 CAD on November 7, 2026; the vendor charges in US dollars.

Two pieces of fine print deserve attention. First, the feature that makes the memory automatic lives in Pro and above: memories your agents can save and recall over MCP. The free tier lets agents search your past sessions, but agents saving memories for later is a paid feature, so judge the free tier on session search, not on self-saving memory. Second, the pricing page disagrees with itself on Team: the plan card says $30 per seat per month plus a $10 monthly base fee, while the comparison table on the same page lists Team at $40 per seat per month. Confirm the real figure before buying for a team.

Enterprise is the only tier with governance: blocking or requiring approval for risky actions, fleet rollout through Jamf, Intune or Fleet, and SSO, RBAC and audit logs. No price is published for it.

Findings: the security half of Beacon

Beacon Cloud reviews your sessions and surfaces three kinds of findings. Security findings flag risky things your agents did, such as a secret pasted into a prompt, a destructive command, an agent bypassing its permissions, or a call to an unknown MCP server. Skill findings spot workflows worth reusing and draft a reusable skill from them. Behavior findings spot habits that cost time or money, such as retry loops, abandoned sessions, approvals you always grant, or expensive models used for simple tasks. Each finding shows what happened, which session it was in, and how to fix it, and one click turns it into a skill or a fix prompt.

This is also where the business model becomes clear. Beacon itself is free telemetry; the enterprise platform built on top of it adds detection, response and governance. That framing matches the funding story: Asymptote Labs announced an oversubscribed $5.5 million seed round on October 20, 2026, led by Tower Research Ventures, with Lytical Ventures, 1984 Ventures and South Park Commons participating. Founders Justin D'Souza and Shukan Shah previously worked at Doppel. Tower Research Ventures says it discovered Asymptote through Beacon's open-source traction and is now evaluating the technology on its own security team's endpoints, which is evaluation by an investor's organization rather than a disclosed commercial deployment.

The company also says Beacon's GitHub stars grew more than 460% in September and that the project reached third on GitHub's global trending repositories that month. Those are company-reported growth figures, not independently verified metrics, and they say nothing about detection accuracy or paying-customer counts. The current star count of 3,363, verified from GitHub on November 7, is the solid number.

Beacon vs Claude Mem: same problem, different scope

This site covered Claude Mem on November 2, a 108,000-star open-source memory tool for AI coding agents. The two tools attack the same problem, but their scope is different. Claude Mem is built for Claude Code; Beacon spans 20+ harnesses, so the knowledge survives switching tools. If you work only in Claude Code, Claude Mem is the simpler pick. If you switch between Claude Code, Cursor and Codex, because one ran out of tokens or a task fits another tool better, Beacon is the one built for that. Our Claude Code explainer and our OpenCode review cover two of the harnesses Beacon connects, and our roundup of the 7 best AI coding agents in 2026 puts the whole category in context.

The catches

Beacon is a strong free tool with a real story behind it, but five catches matter. First, nobody has published numbers on what any of this saves. One independent review puts it plainly: no published benchmarks quantify token or cost savings. Treat the productivity pitch as unproven until someone measures it. Second, the memory that matters most is paywalled: agents saving memories for later needs Pro at $20 a month. Third, harness coverage varies, so check the support matrix for your tools before you assume full capture. Fourth, Beacon records everything agents do, including secrets pasted into prompts, so Local mode is the right default for sensitive work. Fifth, this is a young project: launched in May 2026 with 3,363 stars, it is a long way from the 108,000 stars behind Claude Mem, though 173 releases and a commit as recent as November 7 show active development.

Who should install it

Install Beacon if you switch between two or more coding agents, if you keep re-teaching the same corrections, or if you want a searchable record of what your agents actually did. It is free, open source under the MIT license, and the Local mode keeps your history on your own machine. Wait if you use only Claude Code, if you paste secrets into prompts and will not use Local mode, or if you need the blocking and governance features, which sit in the custom-priced Enterprise tier.

Start on the free tier in Local mode, connect the harnesses you actually use, and let a few sessions accumulate before deciding whether agent-saved memories are worth $20 a month. The memory only compounds if you keep using it, and the review step is what turns a pile of traces into knowledge worth keeping.

Is Beacon really free?

Yes. Beacon Cloud costs $0 forever with no credit card required, and the underlying project is open source under the MIT license. The free tier covers 20+ agents, one timeline, token-spend visibility, 50 findings and session search over MCP. Unlimited findings and agent-saved memories start at $20 a month on Pro.

Does Beacon send my code to the cloud?

Only if you confirm Beacon Cloud during setup. Signing in forwards nothing; the confirm screen states that confirming installs Beacon and starts forwarding your history to Asymptote Labs. Choose Local to keep everything on your machine, and disconnect at any time with beacon endpoint disconnect.

Which coding agents does Beacon support?

The README's support matrix lists 29 local runtimes, including Claude Code, Cursor, Codex, OpenCode, Cline, Devin, Gemini CLI, GitHub Copilot CLI, Kiro, Kimi Code and Qwen Code, plus browser chat on Claude.ai and ChatGPT and cloud agents from Claude Code, Cursor and Devin. Coverage varies by harness, so check the matrix first.

How is Beacon different from Claude Mem?

Claude Mem is built for Claude Code, while Beacon spans 20+ harnesses, so knowledge learned in one tool can be reused in another. Beacon also adds security findings, SIEM forwarding and a paid enterprise governance tier. If you only use Claude Code, Claude Mem is the simpler choice.

Does Beacon replace my agent's built-in memory?

No. Beacon sits across the harness layer, so it adds a shared memory that survives switching tools. Each agent's own memory, context files and project rules still work as before, and Beacon's skills install alongside them.

What does Pro add over the free tier?

Pro at $20 a month adds unlimited findings, full MCP access to sessions, findings and feedback, memories your agents can save and recall over MCP, skills pulled straight from your sessions, a full evidence trail on every finding, real-time alerts and priority support.

Sources

• Beacon official site (beacon.sh): product overview, Jev scoring, install command, Product Hunt badge
• Beacon pricing page (beacon.sh/pricing): Cloud $0, Pro $20/month, Team $30/seat + $10 base, Enterprise custom; findings explained
• Beacon on GitHub (asymptote-labs/agent-beacon): README, MIT license, 3,363 stars, support matrix, install steps, destinations
• Beacon documentation (docs.beacon.sh): installation, runtimes, event schema, MCP, Agent Skills
• Beacon start guide (beacon.sh/start.md): one-line agent setup prompt
• FinSMEs, October 20, 2026: $5.5M seed round, Tower Research Ventures lead
• Automater, October 21, 2026: seed round detail, 460% star growth and GitHub trending claims, founders
• Koda Intelligence, November 4, 2026: independent review, pricing captured November 4, no published savings benchmarks
• AIToolSteps: Product Hunt launch November 2, 2026 and top spot, setup guide, MIT license
• Dev.to, October 25, 2026: Local vs Cloud setup, MCP connect, secrets-in-traces warning
• ChatGate, November 2, 2026: cross-harness reuse, pricing tiers, install options