Notion just gave your agents one brain. Your team needs one too

One afternoon, a developer's agent learns that a geocoding service uses a 503 to say the quota is gone, so a 503 must never be retried. It fixes the bug and moves on.

The next morning, a colleague's agent hits the same service in another repo. It sees a 503 and retries.

The rule is real. It comes from my own tests, and no agent without that lesson in memory ever worked it out by itself. Sonnet, Haiku and Opus all missed it. With a shared memory, every teammate got it right.

This is how teams lose knowledge. Not because their agents are weak, but because what one agent learns stays with one person.

Notion just made the category official

Today Notion announced that ChatGPT, Claude and other agents "can save what they learn to the same memory and use the same skills in Notion." In The Next Era of Notion, Ivan Zhao writes that roughly half the pages in Notion are now created by agents, and that Notion's MCP volume is up roughly 60x this year.

When a company with Notion's reach builds around agent memory, I take that as settled: agents need one.

Credit where it is due: Notion makes setup easy, with one prompt, and a memory is a page people can read and edit. For personal context across your own tools, it is a strong choice.

One brain per person is not a team brain

Look closely at who the memory belongs to. Notion's help page says memories are private to each user. They follow you from Claude to ChatGPT. They do not reach the colleague whose agent is about to retry that 503.

That is the problem I work on with Big Agentic Brain (BAB), the open-source shared memory for coding agents that I am involved with. Its README puts it in one line: "One agent learns, the whole swarm knows."

Notion has a team-level experiment too. Lore is an open-source project that keeps a team's facts in a vault of Notion databases. Here is how the three compare where it matters for a team:

What a team brain needs

Once a whole team writes into one memory, two new problems appear. Several replies to Notion's launch asked the first one: what happens when two agents save opposite lessons?

BAB was built around that. It refuses a memory that duplicates an active one. Colleagues' agents vote on what helped, and every result shows a standing of strong, fair or weak. Two separate flags archive a bad memory, and stale entries are archived too. Maintainers can pin house rules.

The second problem is coordination. BAB gives agents an inbox, messages and reminders, and a key-value store they can use as a lock, so two agents do not work on the same repo at once.

And it stays yours. BAB runs as one container with one SQLite file. It makes no outside calls, needs no LLM keys and works with any MCP client and any model.

The proof

I tested BAB in 1,104 graded runs, described in the README and my first post. On bugs that need a fact only a colleague's memory holds, the share of runs fixed went:

• from 0% to 75% for Claude Haiku 5.5
• from 12% to 79% for Sonnet 5.5
• from 12% to 88% for Opus 5.5

Telling agents to search first took Sonnet to 100%. And the later round showed why local notes are not enough: a Sonnet teammate with only its own notes fixed 8 of 12 related bugs, and every miss was a lesson the shared memory carried.

Give your team one brain

Notion is right that agents need a shared brain. Use it for yourself. For a team of coding agents, the brain has to be shared across people, kept clean as it grows and run on infrastructure you control.

That is what BAB is for. Try it on GitHub. It is free and open source.

If you want help designing memory for your own agents, get in touch at nevolin.be.