Your agent forgets everything when the context window ends. The usual fix is to wire a vector DB, write ingest/retrieve glue, and babysit it. There's a faster path: plug a memory API into
the agent over MCP and let the model call add_memory
/ search_memory
itself — no glue code.
What MCP gives you #
The Model Context Protocol lets an agent discover and call tools. Connect a memory server and the agent gains two tools automatically:
add_memory
— store a factsearch_memory
— recall relevant factsThe model decides when to remember and when to recall, in context — which is exactly where it knows whether something is worth keeping.
Add it to Claude Code #
claude mcp add nexusyn --transport http \
"[https://api.nexusyn.ai/v1/mcp](https://api.nexusyn.ai/v1/mcp)" \
--header "Authorization: Bearer "
That's the whole integration. Your agent now has persistent memory across sessions.
Any MCP client works (e.g. Cursor)
{
"mcpServers": {
"nexusyn": {
"transport": "http",
"url": "[https://api.nexusyn.ai/v1/mcp](https://api.nexusyn.ai/v1/mcp)",
"headers": { "Authorization": "Bearer " }
}
}
}
Why this beats nearest-neighbor search alone
A raw vector store returns similar chunks — but it has no idea which fact is current. When a user corrects something, the old embedding is just as "similar" to the next query. A memory
layer adds the parts that actually make an agent reliable:
Multi-agent, attributed
Several agents can share one memory store. Each write carries its agent, so Claude Code, Cursor and your own app can read each other's memory while you still see who wrote what.
Try it Grab a key, run the one-liner above, and ask your agent to remember something — then start a fresh session and watch it recall. Full guide and the REST equivalent are in the Nexusyn docs
(https://nexusyn.ai/docs/mcp). Disclosure: I work on Nexusyn, a time-aware memory API for AI agents.