Context layer for MCP. A memory server for agents. MCP in. Postgres out.

Rust. pgvector. Self-hosted or managed. Apache-2.0.

Pre-alpha — no code yet.

• First code: 2026-10
• Public beta: 2027-04
• v1.0: usage-driven, not dated

This repository is a placeholder. The landing page and waitlist are the current surface area while the implementation lands.

Agents forget. Context is scattered across vector stores, note databases, in-memory caches, and whatever ad-hoc memory layer the last engineer wrote. Every team building seriously with agents ends up reinventing:

• A vector store wrapper for semantic retrieval
• Session-context scaffolding for cross-turn state
• A per-user / per-tenant isolation layer
• A way to inject all of the above into MCP-speaking agents

Mnemra is that layer, shipped once.

One Rust service. One Postgres + pgvector database. One MCP contract.

        ┌───────────────┐
        │ MCP-aware     │
        │ client        │
        └───────┬───────┘
                │ MCP
                ▼
        ┌───────────────┐
        │ Mnemra server │  ← Rust, Axum, rustls
        └───────┬───────┘
                │ SQL
                ▼
        ┌───────────────┐
        │ PostgreSQL    │  ← pgvector + row-level security
        └───────────────┘

Full spec lands with v0.1.

• Hosted-only product. Self-host is and will remain first-class.
• Vector database. Postgres handles it.
• Frontend. Mnemra is a server.

• Waitlist: mnemra.dev
• Bluesky: @mnemra.dev
• LinkedIn: Mnemra

Not accepting code contributions until v0.1 lands. The surface area needs to be stable before external code is useful. If you want to help shape the direction, join the waitlist — feedback emails go out ahead of each milestone.

Apache-2.0. See LICENSE.

The project reserves the right to relicense future versions. Past contributions remain under their original license.