Memory
Model Context ProtocolA persistent knowledge graph stored in a local JSON file, exposed through entities, relations, and observations
Memory vs AgentDB
A persistent knowledge graph stored in a local JSON file, exposed through entities, relations, and observations On the other side: vector memory with no server to keep running: a single .rvf file holds the embeddings and the cli exposes it over mcp
Installnpx -y @modelcontextprotocol/server-memorynpx -y start
SourceOfficialCommunity
Deciding factorYou want structured facts with visible relationships between them.versus: You want similarity search over notes or documents and can accept opaque ranki
Bottom line
Memory for facts you want to inspect; AgentDB for fuzzy recall you are willing to distrust.
The short answer
Memory is a JSON knowledge graph on disk: entities, relations, observations, easy to read and diff. AgentDB is a vector store in a single file with HNSW search, better at fuzzy recall and worse at explaining why it returned something. Neither is a database you should trust as a system of record.
Facts
| Option | Install | Source | Link |
|---|---|---|---|
| Memory | npx -y @modelcontextprotocol/server-memory | Official | Repository ↗ |
| AgentDB | npx -y start | Community | Repository ↗ |
Pick Memory if
You want structured facts with visible relationships between them.
Pick AgentDB if
You want similarity search over notes or documents and can accept opaque ranking.
What would change our mind
A project going archived, or a maintainer changing the install path, flips this comparison faster than any feature release. Both entries are re-checked weekly, and the date above tells you how much to trust the details.
Read the full entries
More head-to-heads
Filesystem vs Desktop Commander
Gives an agent scoped read, write, and search access to directories you explicitly allow
Comparison
Git vs GitHub
Read-only access to a repository's history: status, diffs, commit logs, and blame
Comparison
Kubernetes MCP vs Kubernetes MCP (kubectl)
Talks to the Kubernetes API directly instead of shelling out to kubectl, so an agent can list contexts, read pod logs, and apply manifests without parsing command output
Comparison
LaunchDarkly MCP vs Unleash MCP
Reads and changes feature flags through the LaunchDarkly API, so an agent can turn a flag off during an incident instead of asking a human to open a dashboard
Comparison
Sentry vs Datadog MCP
Pulls issues, stack traces, and event context into the agent so it can triage production errors with the actual payload in front of it
Comparison
Atlassian Jira MCP vs Linear MCP
Connects to Jira for issue lookup, JQL search, and comments
Comparison