Local-first memory firewall and MCP server for AI agents. Audrey gives coding agents recall, memory capsules, preflight checks, reflexes, validation, tool-trace learning, contradiction handling, and SQLite/sqlite-vec storage.
Overview
Audrey is a local-first memory firewall and MCP server for AI agents.
It helps coding agents remember repository rules, previous failures, validated lessons, contradictions, and tool outcomes before they act again.
Public proof:
- GitHub: https://github.com/Evilander/Audrey
- Benchmark report: https://huggingface.co/spaces/Evilander/audrey-memory-benchmark-report
- Raw benchmark artifacts: https://huggingface.co/datasets/Evilander/audrey-memory-benchmark-artifacts
- AMB provider request: https://github.com/vectorize-io/agent-memory-benchmark/issues/11
Install with:
{
"mcpServers": {
"audrey-memory": {
"command": "npx",
"args": ["-y", "audrey", "mcp"],
"env": {
"AUDREY_EMBEDDING_PROVIDER": "mock"
}
}
}
}
Claim boundary: the public benchmark artifacts are deterministic local regression/performance evidence, not an official AMB, LoCoMo, or LongMemEval leaderboard score yet.
Server Config
{
"mcpServers": {
"audrey-memory": {
"command": "npx",
"args": [
"-y",
"audrey",
"mcp"
],
"env": {
"AUDREY_EMBEDDING_PROVIDER": "mock"
}
}
}
}