post-cortex, created by July Metodiev, is a persistent memory server that extends AI assistants' context across sessions. It converts ephemeral chat interactions into a structured, searchable knowledge base and supplies semantic search and entity catalogs to models that support the Model Context Protocol. Features include local semantic embeddings, multilingual support, non-blocking writes, knowledge graph and high-performance retrieval using lightweight embedding models. Targets developers, power users, and researchers who require private, persistent assistant memory for complex workflows.

Built to convert short chats into reusable long-term knowledge

The tool stores conversation traces and extracts entities into a searchable knowledge graph, making prior decisions and instructions available to later sessions. It targets assistants that implement the Model Context Protocol, acting as a persistent memory layer rather than a transient session buffer. For concrete tasks, it can preserve user preferences, past answers, and named entities so an assistant can reference them across separate chats.

Retrieval favors lightweight embeddings for fast, local lookups

post-cortex performs local semantic search using Model2Vec and BERT-based embeddings, which the developer describes as lightweight options that support quick lookups without heavy hardware. Non-blocking write operations reduce interruption during active conversations, keeping response flow steady while new data indexes in the background. Retrieval accuracy depends on the embedding model choice and the structure of stored traces.

Requires an MCP client and local runtime, with swappable persistence engines

Integration requires an MCP-compliant client and a host runtime supporting Node.js or Bun; configuration typically involves pointing the client to the local server executable. The server supports multiple storage backends, explicitly including RocksDB for file-based performance and SurrealDB for more structured query patterns. This lets implementers pick a storage engine that matches their query complexity and durability needs.

Local-first design protects data but shifts operational responsibility to users

The architecture processes embeddings and entity extraction on the host machine, so no embeddings leave the device and sensitive inputs remain local. Multilingual embedding support broadens language compatibility across global workflows. Community feedback notes the privacy-centric, local-first approach as a distinguishing point, while deployment requires routine local maintenance and platform support across Windows, macOS, and Linux.

Practical choice for technical teams needing private, persistent context

post-cortex is a practical option for developers who need persistent, private assistant memory; expect to operate it inside an MCP workflow and maintain a local runtime. Plan routine export and verification steps if you rely on the memory layer for critical decisions, since long-term continuity depends on local system upkeep and service administration. The tool best fits users prepared to manage local services alongside their assistant stack.