Memory and understanding are different jobs.
A memory layer helps an agent recall what was said. Understanding keeps track of what is still true, who said it, what changed, and what each agent is allowed to see. Both are useful. They answer different questions.
What memory layers do well.
The field is strong and moving fast. This is how the best-known tools describe themselves, in their own papers and docs, checked on 30 September 2026.
mem0
Extracts salient facts from conversations. For each new fact, a language model chooses to add it, update an existing memory, delete one it contradicts, or do nothing. A graph variant links entities.
Mem0 paper, arXiv 2504.19413OpenMemory MCP (by mem0)
A local memory server that runs on your machine and plugs into MCP clients such as Claude Desktop and Cursor. Agents add and search memories through its tools.
mem0 blog: Introducing OpenMemory MCPZep (Graphiti)
A temporal context graph for developers building agents. Facts carry validity windows, are invalidated rather than deleted, can be queried as of any point in time, and trace back to the raw data that produced them.
Graphiti on GitHubLetta
Agents keep memory in blocks that they edit themselves through memory tools. Blocks can be shared between agents, and all state persists in a database.
Letta docs: agent memorySupermemory
A memory API for developers that turns conversations, documents, files, URLs and connector data into extracted facts and per-user profiles, with plugins and MCP for Claude and coding agents.
Supermemory docs
If you are building your own agent and want it to remember its users, these are good choices, and some of them are ahead of us in places. Graphiti’s point-in-time queries, for example, are more complete than what Team0 exposes today.
What understanding adds.
Team0 is built for a different situation: one person, many agents, and a working life that keeps changing underneath them.
- 01
It is built from your work, not from what agents are told
Email, calendar, meetings, documents and the work tools you connect. It arrives already knowing your people and your week, without you writing anything down.
- 02
It knows who said it
Every statement records its source, the exact supporting passage, and whose view it is: yours, someone else’s, or unclear. A stranger’s urgent email never becomes your priority or your instruction.
- 03
Believed, heard and unconfirmed are kept apart
Statements are trusted, evidence or candidate, and the lanes cannot mix. An agent’s own words are kept as evidence under its name, never believed automatically.
- 04
A request is not a promise
A customer asking for June is recorded as a request. Only an agreement makes it a commitment. Something heard stays heard until it is confirmed, and silence never proves a task was done.
- 05
One person, one record
The same email address is the same person across every source. A similar name never merges on its own, a pair you reject is never merged, and every merge can be undone.
- 06
Changes are kept, not overwritten
A newer statement replaces an older one, and the older one stays in history with what changed it. A correction you make once reaches every agent.
- 07
Each agent sees only what you allow
Every agent has its own grant. Its access is applied inside the query, so what it may not see is never retrieved, rather than retrieved and hidden.
- 08
Every read is kept
The exact request an agent made and what Team0 returned are recorded, so you can see what each agent was told.
Side by side.
| A typical memory layer | Team0 | |
|---|---|---|
| What fills it | What agents and conversations send it, plus any data a developer connects | Your connected work: email, calendar, meetings, documents and work tools |
| Unit | Extracted memories or facts, sometimes a graph | Typed statements about people, companies and projects, each with its source and whose view it is |
| When things change | Updated, replaced or deleted; some keep validity windows | Replaced with the older statement kept in history, and the reason recorded |
| Hearsay and guesses | Usually stored alongside everything else | Kept in separate lanes that never become belief on their own |
| Who it serves | One app, or the agents that write to it | Every agent you connect, each with its own access |
The left column is a generalisation. Individual products differ, and the section above describes each in its own words.
Keep reading.
- OverviewThe World Model
- Data modelStatements, two clocks, belief lanes
- Entity resolutionOne person, one record, carefully
- Truth maintenanceHow beliefs are retired, never deleted
- Trust and scopeWho said it, who may see it
- Understanding EngineFrom graph to a read an agent can use
- The 31 problemsEverything you have to solve, in one list
Your agents will change. Your understanding shouldn’t.
Invite-only private beta.