Open-source, cross-harness memory for AI agents | Hacker News Reader

Open-source, cross-harness memory for AI agents

18 pointsshakshukan14 comments

Earlier this year, my co-founder and I found ourselves frequently switching between different coding agents (Claude Code, Codex, Cursor) anytime we ran out of shaky usage limits or a newer, better model came out.

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!

Beacon hooks into 20+ different agent harnesses, captures session telemetry (prompts, tool calls, command executions, etc) in one normalized format, and distills it into common patterns and repeatable workflows with a human-approval step. This compounded knowledge is then accessible to any future agent via skills / MCP.

We believe your agent data is yours to own and should not be locked behind any frontier lab's ecosystem or closed memory infrastructure. We built Beacon to make working across agents way more transparent and efficient.

Try it out and tell us what you think: https://github.com/Asymptote-Labs/agent-beacon

I've found I don't really want my agents to remember very much.

The projects need to remember things, not the agents.

When they're a blank slate coming in, I get more reliable outputs.

I just have one orchestrator/secretary/router with long-term memory, but I don't switch its harness.

I used to be their human clipboard, and one day codex told me it didn't need me..it could call claude by itself, and I found that I could just tell them their thread ID and they could save in their memory and I just ask one to read what the other one did:)

Beacon packages memory into skills, which you can scope to specific repos or projects instead of having it apply generally across any task.

Yeah, I guess now that you mention it, skill files are, in fact, cross-harness memory.

This is intriguing, I think there's probably more compounding value when using across teams / orgs, since I often find myself debugging issues that another engineer's agent already figured out / worked around in a silo, and vice versa.

Agreed, the real unlock is when you can use this as a centralized, "engineering brain". While our OSS is focused on individual developers, our enterprise product provides shared memory & findings across your entire company.

Here's a link to our product page, if it's helpful: https://beacon.sh

Would be worried about an agent remembering too much. a workaround from six months ago doesn’t matter to me anymore. how does a skill get deleted?

You should not use age as a trigger to delete. Instead you should associate each workaround with the bug / version that it was for. Then you can delete this workaround when that bug / version changes.

I am the founder of Substrate, a tool that allows you to see all of the memory and also delete a specific entry. Is your agent allowed to write its own workarounds in your setup?

Hey guys, Justin here, one of the creators behind Beacon! Thank you guys for the love and for getting us to #1 trending on X news!!!

I also built something like this for myself. I call it a markdown file.

Simple markdown files definitely work, but the downside is that they only hold what you actually remember to write down / tell your agents to write down. The useful learnings are usually in the session itself i.e. the 3 approaches the agent tried before one of them worked.

Beacon keeps the full execution trace (prompts, reasoning, tool calls, commands) outside the agent's context and distills lessons from it. Ultimately the agent still only has access to snippets of information (via skill progressive disclousure) but it's drawn from much richer session data rather than what you remembered to note down. Also you don't have to maintain it yourself!

A markdown file works well as long as you don't use a second agent / tool that needs access to this same file. Otherwise, you will find that the copies of the file drift apart.

It's better to store the memory outside of all tools and then connect all of them to it via MCP. Then you have a single store where all of the tools can read from and write to.

This is one of the ideas we used when building up memory in Substrate. How many tools are using this file as memory?

I did too: "keep an obsidian vault of things you need to remember. make a skill for it"

I don't even know how obsidian works.

And I use a harness where the agents can easily look up & search other sessions, or past sessions.

You can add a reference to the source session for each note in the vault. Then you could check if this is true in the actual session instead of trusting what the note says.

This is one of the ideas we used when we were building memory for Substrate.

Does your agent write to the vault automatically?

Reply on news.ycombinator.com