Perspective
Shared memory needs clear boundaries
Durable context helps agents work across time, but access rules and review matter more than simply giving every agent the same vault.
I keep coming back to the same limitation in agent work: useful context disappears too easily. Agents are more useful when they can work from durable context rather than rebuilding an understanding from an empty chat. A well-kept knowledge base can preserve research, decisions, open questions and the working vocabulary around a project.
That does not mean every tool or agent should receive the same access. Shared memory works when the memory has boundaries: public material for public tools, a limited workspace for a specific job, and protected areas for personal, commercial or sensitive information.
Context is a capability
Giving an agent access to a knowledge base changes what it can infer, act on and expose. The questions that matter are practical: which notes does this job require, can it write or only read, what evidence should it preserve, and how can a person inspect the result?
For MCP tools, the same principle applies. A server that bridges notes, files or systems should explain its scope before it is connected. Read-only access, explicit write grants and an audit trail are more valuable than an impressive list of actions with unclear consequences.
Read the full note
This is a short MCP Rack perspective on a broader piece about building durable agent context with Obsidian: I’m Turning My Obsidian Vault Into Shared Memory for AI Agents.
The original was also published on LinkedIn and X.
MCP Rack’s agent access guide describes the public interfaces available here and their limits.