Zengram

Zengram gives AI agents a shared memory across sessions and machines. It's for people using multiple assistants who want a discovery recorded by one agent to be available to the others.

• Typed memories: Keep events as history, update facts by key, track current status and store decisions with their reasoning.
• Cross-agent briefings: Give returning agents updates from other agents and record independent confirmation of a fact as another observation.
• Combined search: Find memories through semantic and full-text search, with optional entity links and reranking to refine retrieval.
• Agent connections: Connect Claude Code, Cursor and Windsurf through the included MCP server, or let other agents access memory through a REST API and Bash adapter.
• Memory cleanup: Redact credentials before storage and use LLM consolidation to merge duplicates, resolve contradictions and identify connections.

Zengram is open source under the MIT license. Its self-hosted memory store uses one Postgres database with pgvector, so it doesn't need a separate vector database service.

Optional reranking works with self-hosted TEI, Infinity or vLLM endpoints, as well as external Jina and Cohere services. Choosing an external reranker sends queries and retrieved memory content to that service for scoring.