Every fact it holds, and every fact it decided not to hand over.

CLI

memvault --data-dir ./my-memory write --namespace project \
  --content "the deploy script lives in ops/deploy.sh"

memvault --data-dir ./my-memory search --namespace project \
  --query "deploy script" --max-tokens 20

memvault --data-dir ./my-memory explain <retrieval_id>
memvault --data-dir ./my-memory verify
memvault --data-dir ./my-memory forget <fact_id> --reason "customer deletion request"

No server and no config. search prints the provenance table above and a retrieval_id; explain reprints that table later from the ledger, and verify answers chain verified from seq 0. The binary builds to target/release/memvault; the commands above assume you put it on your PATH.

MCP

Point the client at memvault-server, with a data directory as its one argument. Save this as .mcp.json in your project root for Claude Code, or merge it into claude_desktop_config.json for Claude Desktop. Any other MCP client takes the same shape: it is a plain stdio server.

{
  "mcpServers": {
    "memvault": {
      "command": "/absolute/path/to/memvault/target/release/memvault-server",
      "args": ["/absolute/path/to/my-memory"]
    }
  }
}

Use absolute paths for both. The server inherits whatever working directory the client happened to launch it from, which is rarely the one you expect.

Restart the client and the agent has seven tools:

On startup the server runs recovery and verifies the chain, logging the result to stderr. Stdout is the protocol channel and carries nothing else.

gRPC

cargo build -p memvault-server --features grpc

MEMVAULT_GRPC_ADDR=127.0.0.1:50051 \
  ./target/debug/memvault-server ./my-memory

Off by default, for multi-process deployments where stdio is not available. The same seven operations, as protobuf messages with the same shapes as the MCP tools’ JSON; the schema is crates/memvault-server/proto/memvault.proto. A default build rejects MEMVAULT_GRPC_ADDR rather than ignoring it, so a half-configured deployment fails at startup instead of quietly serving the wrong thing.

Python

# a wheel for each platform is attached to every release
pip install https://github.com/swap-mitra/memvault/releases/download/v0.2.0/memvault-0.2.0-cp39-abi3-manylinux_2_28_x86_64.whl

import memvault

mv = memvault.MemVault("./my-memory")
fact_id = mv.write("project", "the deploy script lives in ops/deploy.sh")

result = mv.search("project", "deploy script")
for f in result.injected:    # what goes in context, best first
    print(f.fact_id, f.content)
for e in result.candidates:  # why: every candidate considered
    print(e.fact_id, e.outcome, e.final_score, e.token_cost)

mv.verify()  # raises if the chain is broken

In-process, with no subprocess and no event loop. The API is synchronous and every call releases the GIL while the engine works. Explanation carries the same fields as the table above, so provenance does not get thinner just because you came in this way.