Free vs paid memory MCP servers your agent's memory, whose disk?

Memory servers store what you tell your agent — project context, decisions, sometimes trade history or personal notes. basic-memory keeps it in Markdown files you own; superlocalmemory explicitly brands itself local-first and free; in-memoria persists codebase intelligence on your own disk. Privacy and zero marginal cost make local the obvious default for individual use.

These projects also compete on memory science rather than hosting: YourMemory and cognitive-ai-memory both implement decay-and-reinforcement dynamics, where memories fade unless recalled. You get genuinely sophisticated behavior without paying anyone.

Durability and ubiquity. A local SQLite file dies with the laptop, and an agent fleet running across ten machines can't share it. context-sync exists precisely because keeping memory consistent across Claude, Cursor, Windsurf, and Zed is a real unsolved chore — and graph-backed options like mcp-neo4j-memory still require you to run and back up a Neo4j instance.

The moment memory becomes shared infrastructure — team knowledge, multi-agent state, anything that must survive hardware — you're operating a database, with all the costs that implies.

Hosted memory justifies a price when it delivers what local can't: replicated storage, cross-device and cross-agent access, semantic search over large corpora without local GPU or embedding costs, and uptime someone is paid to defend. Specialized domains sharpen the case — a trading agent using something like tradememory-protocol's recall of similar setups is consuming retrieval whose quality directly affects money.

The honest tradeoff is sensitivity: paying for hosted memory means trusting an operator with accumulated context. For some data that's fine; for some it never will be, no matter the SLA.

Memory operations decompose naturally into priced calls: a store, a recall, a semantic search. With x402, each can carry its own price from $0.01, paid by the agent in USDC on Base with ~2-second settlement and an Ed25519-signed receipt — the payment clears before the handler runs, so unpaid calls cost the operator nothing.

Per-call also fits how memory usage actually distributes: heavy during active work, silent for weeks. An agent shouldn't pay rent on memories it isn't touching.

Solo developer, sensitive context, one machine: local-first and free, full stop. Team or fleet, shared knowledge, durability requirements: hosted and paid, with eyes open about what you're entrusting. Many setups end up hybrid — private context local, shared knowledge hosted. All 169 live options are comparable on Loomal's Knowledge & Memory category page.