project
External agent memory backend with a CrossSessionMemoryGuard adapter lacking its own schema attribution layer (KI-11).
No evidence-backed relationships are recorded.
Mem0, Zep, Cognee, Letta, Supermemory and Mnemoverse all ship a knowledge graph, and each one's docs mean something different by a node. Read-time behaviour is the part a buyer can settle from the documentation.
Reality
The MCP schema defaults destructiveHint to true, but only where readOnlyHint is false. A ChatGPT app directory scanner asked for the field anyway on four read-only tools, and the tools came out of it behaving exactly as before.
Reality
The estimate was 30,000 to 40,000 credits per simulation, the gateway needed only three read patterns, and the store already holding tenant rows could serve all three, so tenancy became an argument every query passes.
Reality
A comparison of Mem0, Zep and LangChain's memory classes puts the split at adjudication rather than storage. Mem0 pays two model calls per message to get it.
Reality
A preprint pairs a 'do you know this preference' test with a 'now act on it' test on the same item, and reports a large gap between the two. Health and therapy preferences fare worst.
Reality
A vendor audited 359,388 edges in its own memory store and found feedback-shaped graphs only where its benchmark harness supplied the feedback. Live tenants got one reinforcement per edge.
Reality
A read-only sensor for cross-tenant memory leaks ships with no tagged release, because its own backend cannot enumerate what the retriever exposed. That limitation is the story.
Reality