Best Knowledge & Memory MCP Servers for AI agents.

The contract is simple: store and recall, exposed as tools. Codebase Memory — the category's star leader — builds a knowledge graph over your code in 159 languages with sub-millisecond queries, claiming 99% fewer tokens than re-reading source. basic-memory takes the opposite philosophy: local-first knowledge management in plain Markdown files with bi-directional LLM sync, so your memory store is also a folder you can read, edit, and version yourself.

Graph-backed options sit in between — mcp-neo4j-memory persists agent knowledge as a Neo4j graph, and ori-memory combines identity, a knowledge graph, and search into one memory layer.

A distinctive sub-genre here models memory on how brains work. YourMemory and cognitive-ai-memory implement Ebbinghaus-style forgetting: memories decay with time and strengthen on recall, so what the agent retrieves is weighted toward what actually keeps mattering. shodh-memory layers Hebbian learning over semantic search and knowledge graphs, and neural-memory uses spreading-activation recall across 55 tools.

This isn't academic flourish — naive memory stores accumulate junk until recall returns noise. Decay and reinforcement are pruning strategies, and they're the main axis on which these servers compete.

Decide where the data lives first: local-first servers like basic-memory and superlocalmemory keep everything on your disk, which matters if memories include proprietary code or client details. Then match the structure to the job — a knowledge graph (Codebase Memory, mcp-neo4j-memory) excels at relational queries, markdown vaults excel at human auditability, and specialized stores like tradememory-protocol exist for one domain (trading agents recalling similar setups). Finally check recall quality and token cost: a memory server that returns ten loosely-related blobs costs more than it saves.

The working pattern is write-during, recall-at-start. During a session the agent records decisions, preferences, and facts as they surface; at the start of the next session it queries memory before doing anything else. Tools like in-memoria apply this to codebase intelligence across sessions, and context-sync extends it sideways — syncing context across Claude, Cursor, Continue, Windsurf, and Zed so the memory follows you between clients rather than living in one.

Local-first servers are open source and free to run, and for private memory that's usually the right call. Hosted memory — shared across a team's agents, or too large to keep on a laptop — is a natural per-call service: maintainers can claim their Loomal listing and price store and recall operations from $0.01 in USDC via x402, settled on Base. Loomal's 5% fee on settled transactions is currently waived.