ERA OF MINDS

What is Beacon? Free, self-improving memory for coding agents, and what it really does with your sessions

By Era of Minds ·

Short answer

Beacon (Asymptote-Labs/agent-beacon) is a free, MIT-licensed tool that records what Claude Code, Codex, Cursor, OpenCode and 20+ other coding agents do, then lets you turn the useful parts into project memory that any of those agents can read later. It is real and it works locally. But the memory is not automatic: in the open-source CLI you approve every lesson yourself. And the installer preselects Beacon Cloud, which uploads your last 30 days of agent sessions when you connect. Choose Local if you want nothing to leave your machine.

Era of Minds cover: Self-improving? We checked. Beacon local dashboard screenshot from the agent-beacon README, plus its three plans: Local $0, Cloud $0 with a 30-day session upload on connect, Pro $20 a month.
Figure: Beacon's local dashboard, from the agent-beacon README images (MIT), with three of its plans as listed on beacon.sh/pricing and in the setup guide on October 21, 2026.

What is Beacon memory for coding agents?

Every coding agent forgets. You spend an hour teaching Claude Code that your migrations need a warm cache, then you open Codex the next day and explain it again. Each tool keeps its own history, in its own format.

Beacon sits underneath all of them. It hooks into each harness (or reads its OpenTelemetry output, or polls its logs) and writes one normalized timeline of sessions, prompts, responses, tool calls, commands, file edits, approvals, MCP calls and token use. By default that goes to a local file, ~/.beacon/endpoint/logs/runtime.jsonl.

From that history you can pull out lessons, like "the right way to run a migration here" or "the debugging path that finally fixed the webhook bug," and save them as project memory. Future agents in any harness can then fetch them. The README sums it up in one line: "A problem solved by one agent shouldn't need to be learned from scratch by another."

The facts, checked on GitHub on October 21, 2026 (Dubai time):

  • License and language: MIT, written in Go.
  • Repo: created May 12, 2026. 2,251 stars and 201 forks.
  • Releases: 159 since v0.0.1 on May 13. The latest is v1.3.45, published October 20, 2026 at 06:02 UTC, one of three releases that day. The Linux amd64 tarball is about 88 MB.
  • Coverage: the README's local-agent table lists 29 runtimes, including Claude Code, Codex CLI and Desktop, Cursor, OpenCode, Cline, Gemini CLI, GitHub Copilot CLI, Kiro, Qwen Code and VS Code. Browser chats on Claude.ai and ChatGPT are covered through an extension.
  • Contributors: the top account is co-founder jqdsouza (890 contributions). Second is the claude account with 856, a sign that a lot of the code was written with Claude.

Is Beacon really "self-improving"?

This is where the viral posts and the docs part ways. The launch post says Beacon "continuously distills your sessions into reusable skills." A widely shared repost added that it "runs locally." Both are true in parts, but not together.

In the free, local CLI, the cross-harness memory docs describe the loop as "deliberately review-gated." Here is how it works:

  1. Capture. Hooks record sessions locally. This part is continuous and needs no network.
  2. Score (optional). beacon memory evaluations run sends selected traces to an evaluator: TypeSafe's Jev by default, which needs TYPESAFE_API_KEY, or SubQ. A trace becomes a candidate only if task_success is at least 0.50 and the average score is at least 0.60.
  3. Write the lesson. Per the beacon memory docs, Jev "scores a trace; it does not read it." Its answers are probabilities, so the candidate has no lesson text. You read the trace and write the lesson. Or skip scoring and use beacon memory candidates create with no evaluator at all.
  4. Approve. beacon memory candidates approve saves it to a local memory.db, scoped to that one repository.
  5. Reuse. beacon mcp serve exposes search_memory, get_memory and get_memory_context as read-only MCP tools, or beacon memory skills install writes .agents/skills/<slug>/SKILL.md.

The docs are clear that "nothing in this workflow runs from hooks." Evaluator scores "do not automatically rewrite instructions, install skills, or execute actions." That is a good design choice, because nobody wants an agent quietly rewriting its own instructions. But it means the local version is a memory you curate, not one that improves itself.

The hands-off part, where Beacon notices repeated workflows and drafts a skill for you, lives in Beacon Cloud. The pricing page calls these "findings." Memories agents can save and recall over MCP, and skills pulled straight from your sessions, are listed under Pro.

Does Beacon upload your sessions?

Only if you connect it to Beacon Cloud. But read the setup screen carefully, because Cloud is the default choice.

  • The README says interactive setup "signs in through beacon.sh and preselects Beacon Cloud, with an explicit Local opt-out." Signing in alone forwards nothing. Confirming the Cloud option starts forwarding.
  • Connecting uploads history. The Connect devices page says Beacon sends sessions on disk from the last 30 days so the dashboard isn't empty. Pass --no-backfill to skip that.
  • The default privacy mode is Standard, which, per the Beacon Cloud docs, "sends retained content, such as prompts, responses, tool arguments and results, after local redaction and size limits." Metadata-only (--privacy-mode metadata-only) strips prompts, responses, tool arguments and results, command output and diffs on the device before upload.
  • The copy-paste setup prompt the README suggests ("Read https://beacon.sh/start.md then set me up on Beacon") tells your agent to recommend Beacon Cloud. It also tells the agent to ask you first and never enable a destination you didn't choose.

None of this is hidden. The confirm screen names Asymptote Labs as the operator and states that forwarding will start. It is just easy to click through. If your repos are under NDA, pick Local, or Metadata-only at the very least.

How do you install Beacon for Claude Code, Codex and Cursor?

First the CLI and the endpoint, which does the capturing. From the README:

# macOS
brew update
brew trust asymptote-labs/tap
brew tap asymptote-labs/tap
brew install beacon
beacon endpoint install      # choose Local or Beacon Cloud

# Linux (.deb or .rpm, verified against checksums.txt)
curl -fsSLO https://github.com/asymptote-labs/agent-beacon/releases/latest/download/install.sh
less install.sh
bash install.sh

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

The README suggests reading install.sh before you run it, which is a good habit. For a stronger check that a package came from the repo's release workflow, it also documents gh attestation verify.

Then the Beacon Skills, which let your agent run the memory loop for you:

# Claude Code
/plugin marketplace add asymptote-labs/agent-beacon
/plugin install beacon@beacon

# Codex
codex plugin marketplace add asymptote-labs/agent-beacon
codex plugin add beacon@beacon

# Cursor, OpenCode, Cline and others
npx skills add asymptote-labs/agent-beacon

# Or straight from the CLI (no network needed)
beacon skills install

That gives you five skills: beacon-memory-recall (fetch approved memory before a task), beacon-memory-distill (turn a session into a reviewed lesson), beacon-memory-promote (install a lesson as a project skill), beacon-lens-create (build a dashboard view) and beacon-trace-share (upload a session as a link, only when you ask).

To see what got captured, run beacon traces for a terminal browser or beacon endpoint dashboard for the local web view shown in our cover. Either way, nothing is sent anywhere.

A note on our check: we did not install Beacon on our own machine for this post. It wires hooks into every agent harness it finds, and our box is shared by several agents. So everything above comes from the repo, the docs and the setup guide, read on October 21, 2026.

How much does Beacon cost?

From beacon.sh/pricing and the setup guide, as of October 21, 2026:

  • Open Source, $0: self-hosted capture for 20+ agents, local dashboard, offline threat detection, export to Elastic, S3 or any OTel collector. "No account needed, and nothing leaves your machine."
  • Beacon Cloud, $0: one timeline across machines, 50 findings, session search over MCP, share links.
  • Beacon Cloud Pro, $20 a month ($15 a month billed yearly): unlimited findings, memories agents can save and recall over MCP, skills pulled from your sessions, real-time alerts.
  • Enterprise, custom: blocking or approval for risky agent actions, fleet rollout through Jamf, Intune or Fleet, SSO and audit logs.

So the line "free and open source" is accurate for capture and for hand-curated memory. The most automatic version of memory is the paid one.

Beacon vs CLAUDE.md and AGENTS.md: do you need it?

A reply under the launch post asked the obvious question: how is this different from other memory tools? The simplest comparison is with the files you probably already use.

CLAUDE.md and AGENTS.md are instructions you write by hand. Each harness reads its own convention. They are free, simple and versioned with your code. Beacon's own repo has both.

Supabase's repo is a good example of the hand-written approach: one AGENTS.md, a one-line CLAUDE.md that imports it, and 22 task skills. We looked at it after its CEO joked that 12 Claude agents reverse engineered Supabase.

Beacon starts from the other end: from what agents actually did. It is worth it if:

  • you switch between two or more agents on the same repo and keep re-explaining the same gotchas,
  • you want a searchable record of what agents ran, edited and approved (this is also why security teams use it), or
  • you want token use per session across tools in one place.

If you only use Claude Code on one project, a well-kept CLAUDE.md plus a skill or two may be all you need. Tools like book-to-skill and Ponytail show how far plain skills can go. If your problem is context size rather than memory, Headroom compresses what the agent reads instead.

We have now seen this pattern a few times: a viral claim, a real tool behind it, and fine print that matters. It was 39% fewer tokens but not always a smaller bill with Headroom, and a real 93% gain on paper that lost to "no change" with the Kronos AI trading model. Beacon's fine print is milder. The tool does what it says, as long as you read "self-improving" as "improves when you review it."

A different take on the same problem went viral on October 26: Cindy, an open-source app that runs Claude Code, Codex and Pi side by side and keeps memory and skills when you switch harness mid-task. Beacon captures what your existing tools do; Cindy replaces the front end you run them in.

Who is behind Beacon, and why is it trending on X?

Beacon comes from Asymptote Labs, a New York startup. On October 20, 2026 it announced a $5.5M seed round led by Tower Research Ventures, with Lytical Ventures, 1984 Ventures and South Park Commons. The Next Web reports that co-founders Justin D'Souza (CEO) and Shukan Shah previously worked as engineers at Doppel. The company says Beacon's GitHub stars grew more than 460% in September. Its first market is security: the paid platform watches what agents do inside companies and can block risky actions.

The memory angle is what took off with developers. D'Souza's launch post went up on October 20 at 14:32 UTC (18:32 Dubai). When we checked on the morning of October 21 (Dubai), it had about 435,470 views, 1,309 likes, 1,855 bookmarks, 142 reposts and 252 replies. @kanavtwt's quote post added about 64,790 views and 659 bookmarks, and @akshay_pachaar posted a walkthrough with about 17,090 views and 252 bookmarks.

Nearly 1,900 bookmarks on one post means a lot of people plan to try it. If that's you: install it, pick Local first, and run it for a week before you decide whether Cloud is worth it.

Key facts

  • Tool: Beacon, an open-source, cross-harness session capture and reviewed-memory layer for coding agents.
  • Repo: Asymptote-Labs/agent-beacon. MIT, Go. Created May 12, 2026. 2,251 stars / 201 forks (checked Oct 21, 2026).
  • Latest release: v1.3.45 (Oct 20, 2026). 159 releases since May 13.
  • Agents: Claude Code, Codex, Cursor, OpenCode, Cline, Gemini CLI, Copilot CLI and more (29 local runtimes in the README table).
  • Memory: review-gated. Optional Jev or SubQ scoring (needs an API key), candidate needs task_success ≥ 0.50 and score ≥ 0.60, you write and approve the lesson. Served via MCP or as SKILL.md.
  • Cloud: preselected in interactive setup; connecting uploads the last 30 days of sessions unless --no-backfill. Standard mode sends prompts and responses after redaction; Metadata-only strips them.
  • Pricing: Open Source $0, Cloud $0 (50 findings), Pro $20/mo or $15/mo yearly, Enterprise custom.
  • Company: Asymptote Labs, New York. $5.5M seed led by Tower Research Ventures (Oct 20, 2026).
  • X buzz: @jqdsouza launch post ~435,470 views / 1,309 likes / 1,855 bookmarks (checked Oct 21).

FAQ

What is Beacon for coding agents?

Beacon (GitHub: Asymptote-Labs/agent-beacon) is a free, MIT-licensed tool from Asymptote Labs that records what your coding agents do across Claude Code, Codex, Cursor, OpenCode and 20+ other harnesses. It saves prompts, tool calls, commands, file edits, approvals and token use as local JSONL. You can then turn selected sessions into reviewed project memory that any agent can read back through MCP or as an Agent Skill.

Is Beacon really self-improving memory?

Not on its own. In the open-source CLI, memory is deliberately review-gated. You pick sessions, an optional evaluator scores them, and a candidate only qualifies if task_success is at least 0.50 and the average score is at least 0.60. The evaluator returns probabilities, not lesson text, so you write the lesson and approve it. Nothing in the memory workflow runs from hooks. The automatic skill drafting the launch post describes is part of Beacon Cloud findings.

Is Beacon free, and does it run locally?

Yes to both, with one catch. The open-source version and the basic Beacon Cloud plan are $0. Beacon records locally by default to ~/.beacon/endpoint/logs/runtime.jsonl. But the interactive installer preselects Beacon Cloud, and connecting a device uploads the last 30 days of agent sessions on that machine unless you pass --no-backfill. Pick Local during setup if you want nothing uploaded. Beacon Cloud Pro, which adds memories and skills over MCP, costs $20 a month or $15 a month billed yearly.

How do I install Beacon for Claude Code?

On macOS run brew trust asymptote-labs/tap, brew tap asymptote-labs/tap and brew install beacon, then beacon endpoint install and choose Local or Cloud. On Linux use the release .deb, .rpm or install.sh (it checks the package against checksums.txt). On Windows use the x64 MSI. For the memory skills inside Claude Code, run /plugin marketplace add asymptote-labs/agent-beacon and then /plugin install beacon@beacon, or run beacon skills install.

How is Beacon different from CLAUDE.md or AGENTS.md?

CLAUDE.md and AGENTS.md are instruction files you write by hand, and each harness reads its own. Beacon builds memory from recorded sessions instead. An approved lesson can be served to any MCP-capable agent through beacon mcp serve (search_memory, get_memory, get_memory_context) or written out as .agents/skills/<slug>/SKILL.md, so a fix found in Cursor can reach Codex or Claude Code. You still review every lesson.

Do I need an API key to use Beacon memory?

Only for automatic scoring. beacon memory evaluations run calls TypeSafe's Jev evaluator by default and needs TYPESAFE_API_KEY (or SubQ with SUBQ_API_KEY). You can skip the evaluator completely: beacon memory candidates create lets you write a lesson from a trace yourself, with no network call. Capture, review, MCP retrieval and skill install need no key.

Who makes Beacon?

Asymptote Labs, a New York startup co-founded by Justin D'Souza (CEO) and Shukan Shah. On October 20, 2026 it announced a $5.5M seed round led by Tower Research Ventures, with Lytical Ventures, 1984 Ventures and South Park Commons. Beacon is its free core product; the paid platform adds org-wide findings and blocking of risky agent actions.

Sources

Keep reading