Most agent architectures treat memory as a retrieval problem. Multiple agents share a vector store and rely on metadata filtering, routing logic, or prompt-level rules to control what each agent can see.
In practice, this becomes hard to reason about as systems grow.
Moreover, I found that memory in agent systems is not just storage. It also becomes a coordination mechanism and a governance surface for knowledge written by autonomous processes.
CtxVault explores a different abstraction.
Memory is organized into independent knowledge vaults with separate retrieval paths. Vaults act as controlled knowledge scopes that agents can attach to at runtime.
The server exposes vault names as part of the API design. This means isolation depends on how agents are implemented, similar to how system-level primitives provide capabilities without enforcing policy.
The goal is to provide a controllable, semantic, and flexible memory layer that can be used for both shared knowledge and isolated workflows, depending on how systems are built on top of it.
Vaults can be inspected and managed manually. Agents can persist semantic memory across sessions using local embedding and vector search pipelines.
The system runs fully local using FastAPI as a control layer.
I am particularly curious about real-world experience with long-term agent memory. When building production systems, do you find yourself relying more on architectural separation of memory domains, or on smarter retrieval/routing strategies?