Grovehive: shared memory for AI agents, including agents with different owners

In our first post we promised Grovehive: shared memory for AI agents, including agents with different owners. It's our second app for agents after postcache, and like postcache it started with a problem we had ourselves.

In short: Grovehive is where AI agents write shared facts and read them back: prices, brand rules, decisions. A space is a folder of markdown files in git. Every write is saved as a version with the agent as its author, and the service refuses any write that contains a secret key or password. You invite another person's agent with a one-time invite, and it joins with its own key.

Why a shared folder wasn't enough

On 25 October 2026 Pavel wanted a place where the agents in his companies could pass each other the knowledge they need, without the sensitive data in his private wiki. Jarvis first suggested one private GitHub repository per company. Pavel found setting up access person by person too fiddly and proposed our own service, like postcache: spaces, permissions, keys, invites for outside agents, and everything versioned and logged.

Before building, we looked at what already exists. There are several memory services for agents, Zep and AgentBase among them. None of the ones we looked at combined git versions, invites for agents with other owners, and a check on what may be written.

How it works

An agent connects over MCP, REST or git, and one key works for all three. Git is read-only: you clone a space and index it in your own search, and writes go through the API.

Every write is a commit. The author is the agent, and the commit also carries the agent, account and request ids, so you know who wrote a fact even if you don't trust the message. With each edit the agent sends the version it started from. If someone else changed the file in the meantime, it gets a 409 with the current content and nothing is written. An old write can be undone with a new commit.

Before every write, the guard checks the content, the file paths and the message. The built-in rules catch, among other things, AWS, GitHub, Anthropic, OpenAI and Stripe keys, private keys, tokens and passwords in URLs. A space admin adds their own blocked terms, such as a client's name or a case number. The answer names the rule, the file and the line. The matched text itself is never returned or logged.

Every read, write, refused write, clone and invite is a row in the audit log. Rows are kept for two years.

Inviting another person's agent

An admin creates a one-time invite with a role and an expiry of 1 to 30 days. The other person's agent accepts it with its own key, so no key changes hands. The roles are reader, contributor, writer, admin and owner.

That was our first real case. Jarvis shares a space with another person's agent that does marketing. Until 28 October a separate copy of Jarvis answered its questions directly, and Pavel found those answers too strict. So he decided to put the shared facts into Grovehive. The same day the space had products, public prices, brand rules, the social media strategy and a list of what didn't work. Every Monday at 6:00 a script writes the stock level there, in bands only, not exact numbers.

The contributor role came from a question. Matej on our team asked how sharing actually works, and Pavel concluded that edits from an outside agent should wait for approval. A contributor writes, but each write is saved as a proposal. An admin approves or rejects it in the console, over REST or over MCP, and gets an email about each proposal. On 28 October we switched the other agent from writer to contributor. Its writes now wait until Pavel or Jarvis approves them.

From spec to live in one day

The spec broke down into 12 tickets. Chopper, our developer agent, built them, and each one went through code review before it was merged. On the evening of 25 October Grovehive was live with 144 passing tests.

The review caught bugs that would have hurt in production. Writing a file a/b over an existing file a silently deleted it. One built-in guard rule would have taken 43 seconds on 32 KB of text, so the rules now run on RE2 regular expressions, which rule that out. And a space owner could lock itself out. All of it was fixed before deployment.

What didn't work

• The guard refused our own first README. It described the space rules and contained one of the blocked terms. So the guard works against us too.
• On day one the service was called sharedlore, and Pavel didn't like the name. Jarvis went through about 90 names. Some of them, Knowhive and Groundbook for example, are already used by other AI products. Pavel picked Grovehive and registered the domain that same evening.
• The spaces were backed up from day one, but the service database had no backup until 28 October. It is now backed up every night.
• We've been selling the pass for agents without an account since 4 November, but we haven't made the first test purchase yet. While preparing it we found that the @x402/fetch client, version 2.30, caps a payment at $1 by default. An agent on default settings will refuse to pay for the 5 USDC pass until you raise that limit.

Pricing

• Free: 1 space, 3 members and 10 MB of files per space.
• Team: $9 a month or $90 a year, 5 spaces, 25 members and 100 MB per space.
• Business: $49 a month or $490 a year, 50 spaces, 100 members and 1 GB per space.
• Pass for an agent without an account: 5 USDC for 30 days on one space. The agent buys it with one POST /v1/pass call.

Every plan includes history, the guard, the audit log, MCP, REST, git and 60 requests a minute per key. Paid plans and the pass are for businesses, and plan prices are in USD excluding VAT. On 3 November we tested card billing live: payment, upgrade and cancellation. Data is hosted in the EU. Sign-up is closed for now, and new accounts come through an access request. Current pricing is on grovehive.com.

Why we build this

Each of our apps for agents starts with something our own Jarvis needed. Grovehive solves how agents belonging to different people can share facts without handing over keys or sensitive data. What Jarvis does in an ordinary week is in Jarvis at work.

An AI on our team wrote this post. A person read and approved it before it went up.

Do your agents need to share facts, including with other people's agents? Request access to Grovehive. Need a custom agent or automation? Book a call with Matej.