ArgosBrain · For MCP developers

A memory MCP server
that doesn't call an LLM
to remember.

Drop-in Model Context Protocol server. Seven deterministic tools. Rust core, stdio transport, zero-panic in production. Persistent memory for agents you build — without paying tokens on the retrieval path.

View on GitHub Install FAQ
01The problem

Most "memory MCP servers" still call an LLM to read.

If you've surveyed the MCP memory landscape you've seen the same pattern: the tool is named recall, but under the hood it embeds the query, calls a cloud model to summarize, or proxies through OpenAI. Every turn your agent takes, the memory layer taxes it. The token bill compounds.

The other half of the landscape is flat-file: CLAUDE.md, markdown drops, prose injected into the prompt. That's fine for "the user prefers async/await," useless for "enumerate every caller of DatabaseClient.connect."

If you're shipping an MCP-based agent to real users, neither is acceptable infrastructure.

02Seven tools. All deterministic.

What ArgosBrain exposes over MCP.

recall
Association-based graph traversal. Returns raw code chunks ranked by activation.
search
Hybrid keyword + HNSW vector search. Sub-millisecond at 100K+ nodes.
symbol_exists
Precision Layer. Boolean: is this name defined in the current project scope?
resolve_member
Does foo.bar exist? Answers member-access questions deterministically.
list_symbols
Enumerate symbols by kind, module, or file. Zero LLM cost.
naming_convention
Learn the project's naming style so new names fit.
introspect
What does the brain know? Zone distribution, node count, confidence summary.

Plus the write path: remember, learn, verify, connect, strengthen, forget, consolidate. Hebbian reinforcement and decay run in a background tokio task.

03Wire it up

Claude Desktop / Cursor / custom agent.

{
  "mcpServers": {
    "argosbrain": {
      "command": "argosbrain-mcp",
      "args": ["--project", "/path/to/repo"]
    }
  }
}

Stdio transport. No network. No API key. The binary starts in under 100ms and holds the full symbol graph in ~50MB RAM for typical repos.

If you're building a custom agent with the Claude Agent SDK, TypeScript MCP SDK, or Python SDK — the server speaks standard MCP. No vendor lock-in. No bespoke protocol.

04Why Rust, why in-process

The architecture is the product.

  • In-process Rust — no external DB (no Neo4j, no Redis, no Postgres). Ingestion, retrieval, decay — all in one binary.
  • petgraph + HNSW + keyword — hybrid retrieval that beats pure vector on code and pure keyword on natural language.
  • Zero panics in the stdio path. If the brain hits an unhealthy state it goes into lameduck — the client keeps working, the brain self-recovers.
  • File-hash invalidation — refactors don't poison the graph. Stale chunks are detected and replaced automatically.
  • Project-scoped — multi-project isolation by design. Memory from repo A cannot leak into repo B.

The benchmark (LongMemCode) is MIT-licensed, and every number on this page is reproducible on your laptop using it. The engine itself is commercial — the source is not published, but adapter stubs for competitors in the benchmark repo let anyone run the comparison.

05Next

Clone it, install it, or read the architecture docs.

github.com/CataDef/neurogenesis → · One-line install · Independent verdict · Compare vs other MCP memory servers