Founder, Engraphis
What should an MCP memory server remember between coding sessions?
I keep seeing the same failure: the agent can solve a task in one session, then starts the next session without the decisions that shaped the code.
The questions I use before trusting a memory server:
- What is the claim scoped to: repo, service, user, or session?
- When did it stop being true, and what replaced it?
- Which record supported the answer?
- Can it abstain when the record is missing?
If it can answer those four, it is a memory system. If it only returns similar passages, it is a search box with extra steps.
I wrote the 12-question version here: engraphis.com/agent-memory-checklist
Disclosure: I maintain Engraphis, a local-first memory engine for coding agents.






Comments
0 comments