Engram

Engram is a cited, durable memory system designed for AI agents, providing explainable recall and persistent storage of conversations, facts, and decisions. It enables agents to retrieve information with citations to original source memories, supporting transparency and traceability in automated reasoning. The platform operates over MCP (Memory Control Protocol) and is accessible via both MCP and HTTP, allowing integration with a range of agent frameworks and popular IDEs. Users can bring their own language model by supplying an API key from providers such as OpenAI, Anthropic, Groq, Together, or Fireworks, with all inference routed through the user's chosen provider.

The retrieval engine in Engram runs three methods in parallel: BM25 for exact lexical recall, vector search for paraphrased matches, and a knowledge graph for relational queries. Results from these engines are fused using reciprocal rank fusion, reranked with a cross-encoder, and returned alongside an explanation trace that details which engines contributed, the graph facts involved, and the scoring of each memory. This approach is designed to provide explainable memory, where every answer is accompanied by a trace showing the reasoning process and supporting evidence. Engram also includes background agents—such as Watchdog, Janitor, and Bucket Profiler—that manage retention, cleanup, and allow rollback of mutations within 90 days.

Integration is streamlined, with one-line installation available for Claude Code's plugin marketplace, as well as manual CLI setup for direct MCP server connection. The system works with agent runtimes and IDEs like Claude Code, Cursor, and ChatGPT, and offers over 55 integrations. Security measures include encryption of user model keys with AES-256-GCM, and Engram does not log prompt bodies or use customer content for training.

Engram offers a free tier that includes 10,000 memories and 50,000 retrievals per month, with no credit card required. Paid plans are available, including Indie and Team tiers with increased memory and retrieval limits, and a custom Enterprise plan offering unlimited capacity and managed inference. The composer prompt used in Engram's benchmark is MIT-licensed, and the platform distinguishes itself by never charging per-token fees, as billing is handled directly by the user's LLM provider.