· 2 min
The small reader missed facts that were already in the prompt
I wrote a preprint on this. On one development split, most of the misses happened with the evidence already loaded. I am not claiming a win over Mem0.
blog
Blog
Direct answers first. Then the longer explanation.
· 2 min
How I actually built the memory
Sources come in. Notes keep the date and who said it. A newer decision marks the older one. An MCP client receives the current context.
· 1 min
The agent was useful the second time, not the first
The first session is a briefing. The second session is where memory either shows up or you paste the same thread again.
· 1 min
The old decision was still in the notes
On 1 October we parked SSO. On 2 October we shipped it with the Team plan. Both notes exist. Only one is current.
· 1 min
I did not want a store I have to query
A store waits for a query. I wanted the agent to receive the current context when it starts work.
· 2 min
Mem0, Zep, Letta, and what I am not claiming
They are memory systems you query from your own agent. I give the current context to the agent. There is no head-to-head score on this page.
· 1 min
I use more than one AI in a day
The memory should not live inside one chat product. I connect the tools once and point each MCP client at the same server.
· 1 min
I would rather the agent say it does not know
A fluent sentence with no note behind it is the bug I am trying not to ship.
· 1 min
MCP is the socket, not the memory
MCP is how a client calls a tool. Archilas is the tool that hands back the current context.
· 1 min
The context died when I closed the tab
The model did not forget. The session ended. The next tool opened with an empty prompt.
· 1 min
Search found the paragraph. It did not know which decision won
RAG is how I pull a passage. It is not how I know the October decision replaced the September one.