Compare · Memory
Memory vs Memory Bank MCP Server
Memory exposes 9 tools and adds about 949 tokens to every model turn. Memory Bank exposes 5 tools and adds about 323. Memory costs 2.9× as much context. Measured by connecting to each server and counting its tool list, not from the README.
 Memorymodelcontextprotocol |  Memory Bankalioshr | |
|---|---|---|
| Context tax | ~949 tok/turn | ~323 tok/turn(leaner) |
| Tools | 9 | 5 |
| Cost per 1,000 turns (Claude Sonnet 5.5, uncached) | $1.90 | $0.65 |
| Security score | 95/100 | 87/100 |
| Cold start | 791 ms | 219 ms |
| Runs | Local (stdio) | Local (stdio) |
| Needs keys | None declared | None declared |
| Freshness | Active | Abandoned |
| GitHub stars | 91k | 921 |
| License | — | MIT |
| Official MCP registry | No | No |
Install in Claude Code
Memory
claude mcp add memory -- npx -y @modelcontextprotocol/server-memory
Memory Bank
claude mcp add memory-bank -- npx -y @allpepper/memory-bank-mcp
Tools
By name, from each server's live tool list. The same name can still behave differently.
Only in Memory · 9
- create_entities
- create_relations
- add_observations
- delete_entities
- delete_observations
- delete_relations
- read_graph
- search_nodes
- open_nodes
In both · 0
- none
Only in Memory Bank · 5
- list_projects
- list_project_files
- memory_bank_read
- memory_bank_write
- memory_bank_update
Questions
- Which uses fewer tokens, Memory or Memory Bank?
- Memory Bank: its tool definitions take about 323 tokens per turn against 949 for Memory (2.9× more). Tool definitions are resent on every model turn.
- Do Memory and Memory Bank have the same tools?
- No tool names overlap: Memory has 9 tools and Memory Bank has 5.
- Are Memory and Memory Bank hosted or local?
- Memory: local (stdio). Memory Bank: local (stdio).
Token counts use the o200k tokenizer on each server's tool definitions as an agent receives them; clients that load tools lazily pay less. Prices are Claude Sonnet 5.5 input at $2/MTok. Want to check one yourself? Run npx mcpskillhub test <package> — how it works.