Mind · MCP Bridge

Memory for the Agent You Already Use.

Fourteen tools. Five lines of config. Your Claude Code, Cursor, or Codex CLI reading from the same compounding graph.

Direct answer

Anvaya’s MCP memory server gives any MCP-compatible agent persistent memory: Claude Code, Cursor, Codex CLI, Windsurf, and more connect with five lines of .mcp.json and get 14 tools, mind_query (PackedContext under a 4096-token budget), mind_search, mind_get_node, mind_record_change, mind_feedback, mind_stats, session start/end, thread and delta retrieval, proactive candidates, intent, retro-promotion, and opt-in federation. The server is your local Mind daemon (SQLite + HNSW under .anvaya/mind/), so nothing is hosted and nothing leaves your machine. Anvaya CLI itself uses faster native transports (in-process Direct under 0.5ms, Unix-socket IPC around 2ms); MCP at ~10ms is the universal bridge. One graph, every harness: record a fix in the terminal, recall it in the IDE.

The 14 Tools

A Memory API, Not a File.

Reads and writes, the agent accumulates knowledge, not just consumes it.

01mind_query (budget 4096)The main read path: natural-language query returns a PackedContext (critical nodes, patterns, fixes, concepts) capped at a 4096-token budget.

02mind_search · mind_get_nodeHybrid semantic + keyword search across the graph; full node fetch with links, confidence, and utility history.

03mind_record_change · mind_feedbackWrite paths for agents: record decisions, fixes, and turn feedback so future sessions (any client) benefit.

04mind_stats · threads · deltaGraph stats, causal thread retrieval, and delta queries for what changed since a session.

05mind_proactive · mind_intentIntent-aware proactive candidates and the session's inferred intent, memory that can volunteer under an EVOI gate.

06sessions · retro_promote · federateExplicit session start/end, knowledge recovery, and opt-in cross-project export/import.

Setup

Five Lines.

Drop in project root as .mcp.json, restart the client, query.

01Anvaya's MCP server: over stdioThe server speaks MCP stdio from your local memory daemon, no ports, no tokens, no accounts.

02Same graph everywhereSessions from CLI, IDE, and CI land in one store with per-project isolation and cross-project federation when you opt in.

03Inspectable alwaysmind_stats shows nodes, sessions, and injection utility; the SQLite file opens in any browser.

Questions

Asked About the MCP Server.

How do I add memory to Claude Code?

Point Claude Code at Anvaya's MCP memory server: five lines of .mcp.json in your project root. Claude Code then gets 14 memory tools (query, search, record, feedback, stats, sessions, threads, proactive, intent, federation) backed by your local graph. No account, no hosted service; memory stays under .anvaya/mind/.

Does it work with Cursor, Codex, or other MCP clients?

Yes, any MCP-compatible client speaks the same protocol. Anvaya CLI uses faster native transports (Direct <0.5ms, IPC ~2ms); MCP (~10ms) is the universal bridge for third-party harnesses.

What does a budgeted query mean?

mind_query takes a token_budget (default 4096) and the server packs the best context that fits: top critical nodes first, then patterns, fixes, concepts. Your agent gets high-signal memory at a fixed, predictable cost per call.

Is my code sent anywhere?

No. The MCP server is your local daemon over stdio/Unix socket. Queries never leave your machine unless you independently configure a cloud LLM provider for the agent itself.

Run Agents That Fit On Your Laptop.

25.6 MB median RSS. 25 agents ran in parallel on a Core 2 Duo with 4 GB RAM. Hundreds on your machine. Zero cloud required on the Ollama path.

Get Started →Read the FAQ

Requires Rust/cargo to build from source. Linux and macOS today, Windows not yet supported. Pre-1.0, public beta. Pricing TBD.