Problem
MenteDB's cognition-aware storage is compelling for single-agent deployments, but multi-agent systems require memory isolation between agents that share the same DB instance. Without namespace scoping, agent A's knowledge graph edges and HNSW vectors are visible to and retrievable by agent B, even when their domains are completely separate.
In practice: a coding agent and a research agent sharing a MenteDB instance would contaminate each other's speculative context pre-assembly, since the vector retrieval has no way to know which memories belong to which agent.
Proposed addition
A lightweight tenant/namespace concept attached to memory nodes and edges:
- A
agent_id or namespace field on all stored entries (graph nodes, vector embeddings, WAL entries)
- Retrieval APIs accept an optional namespace filter, defaulting to global when absent
- Cross-namespace reads available as an explicit, privileged operation (not the default)
This would make MenteDB usable in team-agent deployments (e.g. an engineering agent + a QA agent + a reviewer agent all backed by the same MenteDB instance) without memory bleed.
Why this matters for the project
The 'cognition-aware' tagline implies MenteDB models how agents think -- but how they think is identity-specific. Without isolation, the cognition layer is actually shared across identities, which is the opposite of what a persona-driven multi-agent system needs.
Problem
MenteDB's cognition-aware storage is compelling for single-agent deployments, but multi-agent systems require memory isolation between agents that share the same DB instance. Without namespace scoping, agent A's knowledge graph edges and HNSW vectors are visible to and retrievable by agent B, even when their domains are completely separate.
In practice: a coding agent and a research agent sharing a MenteDB instance would contaminate each other's speculative context pre-assembly, since the vector retrieval has no way to know which memories belong to which agent.
Proposed addition
A lightweight tenant/namespace concept attached to memory nodes and edges:
agent_idornamespacefield on all stored entries (graph nodes, vector embeddings, WAL entries)This would make MenteDB usable in team-agent deployments (e.g. an engineering agent + a QA agent + a reviewer agent all backed by the same MenteDB instance) without memory bleed.
Why this matters for the project
The 'cognition-aware' tagline implies MenteDB models how agents think -- but how they think is identity-specific. Without isolation, the cognition layer is actually shared across identities, which is the opposite of what a persona-driven multi-agent system needs.