SciblyGuide · · 2 min read
AI agent memory for teams: who owns what it learns
CLAUDE.md, auto memory and Mem0 remember a lot, but for one person. How an agent's memory becomes the team's knowledge, with each reader's permissions.
Niclas GregorCo-founder of Scibly
The key points at a glance
- Agent memory is what an AI agent keeps beyond one session: instructions, corrections and what it has learned.
- Almost every such memory belongs to one person and one tool. So everyone teaches their own agent everything, and what Claude Code knows, Cursor doesn't.
- Sharing everything doesn't work either: sessions hold half-finished work, personal things and sometimes passwords.
- Scibly becomes one memory that all of a person's agents share, whatever tool they run in. Everything arrives private; with the share switch on, colleagues see the facts from it, never the text.
What is agent memory?
Agent memory is everything an AI agent keeps beyond one session: instructions, corrections, preferences and what it learned in earlier sessions. In Claude Code these are the CLAUDE.md files you write and the auto memory Claude writes itself (docs). If you build your own agent, you use a library such as Mem0.
Who owns an agent's memory?
The person who uses the agent, and the tool it runs in:
- Claude Code. The
CLAUDE.mdin your home folder applies only to you, and auto memory lives on your machine. Only theCLAUDE.mdin the repository is shared, through Git (docs). - Claude. Memory belongs to each person alone, and not even the organisation's owners see it (help).
- Mem0. Memories are kept per person, agent and session (docs). Who reads what is up to your application.
For one person, that is right. In a team of three, it looks like this:
./CLAUDE.mdShared through Git, so only with those who have the repository.
What is missing when every agent has its own memory?
- Everyone teaches their own agent. Jonas tells his agent that staging runs on the new runner. Mia's agent keeps suggesting the old one.
- It is tied to one tool. What Claude Code knows, Cursor doesn't.
- It stays with the person. When someone changes teams, the memory stays on their machine, out of the team's reach.
Sharing everything doesn't work either: an agent's sessions hold half-finished work, personal things and sometimes passwords. A memory for the team needs one shared place where each person reads only what they may.
How does Scibly turn it into the team's knowledge?
Scibly does not replace the memory of Claude or Claude Code. It becomes one memory that all of a person's agents share, whatever tool they run in, and from there one for the team:
- Bring what is already there. Once you agree, your agent sends your sessions, memory files and skills (saved instructions for the agent) to Scibly. It leaves out what tools printed and
.envfiles with passwords. - Everything arrives private. No one else sees it. Your agent then finds it in every tool that speaks MCP, Cursor and ChatGPT too.
- Share without sharing the text. Each person has a switch per source, such as their agents' sessions. When it is on, colleagues see the facts Scibly read from it, with the person as their source, but never the text itself. The switch stays off until the person turns it on; no one else can do that for them.
Start with what your agents already know today. Give your agent this task:
Read every memory and instruction file you load in this project and in my user configuration, such as CLAUDE.md, AGENTS.md and your memory files. Sort each line into three groups: only for me, for the whole team, and lines with keys or personal data that belong in no file. For the team lines, tell me who on the team doesn't get them today.
More on how engineering teams use Scibly: Scibly for engineering teams.
