Best memory MCP servers
We installed and handshake-tested 11 tested agent tools and MCP servers for memory. Short answer: Pinecone (official) (100/100). This list includes paid and keyed servers; servers that need a key are labelled, and we have not tested any of them with real credentials.
Last tested 2026-10-22
How we test: methodology. Not tested by us: security, output quality, and behaviour with real credentials.
Ranked by test score
| Server | Score | Status | Tools | Schema tokens (est.) | Functional tasks | GitHub stars |
|---|---|---|---|---|---|---|
Pinecone (official)@pinecone-database/mcp | 100 | working | 9 | ~5,696 | store entity: not run (-) | 72 |
Memory / knowledge graph (reference)@modelcontextprotocol/server-memory | 93 | working | 9 | ~925 | store entity: pass (create_entities) retrieve entity: pass (read_graph) | 91,469 |
Basic Memorybasic-memory | 93 | working | 21 | ~7,293 | store entity: pass (write_note) retrieve entity: pass (search_notes) | 4,310 |
Qdrant (official)mcp-server-qdrant | 90 | working | 2 | ~232 | store entity: not run (qdrant-store) retrieve entity: not run (qdrant-find) | 1,552 |
Knowledge Graph memorymcp-knowledge-graph | 80 | working | 10 | ~3,610 | store entity: fail (aim_memory_store) retrieve entity: fail (aim_memory_read_all) | 886 |
Knowledge Graph (itseasy21)@itseasy21/mcp-knowledge-graph | 75 | working | 11 | ~947 | store entity: pass (create_entities) retrieve entity: pass (read_graph) | Not measured |
Neo4j Memory (official)mcp-neo4j-memory | 60 | failed | Not measured | Not measured | Not measured | 989 |
lbrain agent memorylbrain-coding-agents | 55 | failed | Not measured | Not measured | Not measured | 0 |
Chroma (official)chroma-mcp | 40 | failed | Not measured | Not measured | Not measured | 601 |
Memory (Python)mcp-server-memory | 35 | failed | Not measured | Not measured | Not measured | Not measured |
Memory Keepermcp-memory-keeper | 25 | install-failed | Not measured | Not measured | Not measured | 136 |
Ranked by score, then GitHub stars. Failed = did not complete the MCP handshake without credentials or a backing service.
Small-model runs for Memory / knowledge graph (reference) (directional, 3 runs each): qwen2.5-3b-instruct-q4km completed 100% of tasks; qwen2.5-1.5b-instruct-q4km completed 67% of tasks. MCP with local models.
FAQ
What is the best memory MCP server?
In our 2026-10-22 tests, Pinecone (official) (@pinecone-database/mcp) ranked first with 100/100 and 9 tools listed. Ranking uses our test score, then GitHub stars as a tie-break.
Which memory MCP server uses the least context?
Qdrant (official) lists 2 tools at about 232 schema tokens, the lowest among maintained servers here that we measured.
Which memory MCP servers are deprecated?
None of the 11 we tested, as of 2026-10-22.