Plenty of memory servers return an answer; this one returns the memories and graph edges that produced it, which is what makes a wrong answer debuggable. Organising by bucket — work, personal, one per project — is not decoration either: retrieval is scoped to a bucket, and the MCP tool takes one at a time.
A hosted memory service reached over SSE. Memories go into named buckets; retrieval fuses BM25, vector search and a knowledge graph, reranks the result, and returns an answer with the memories that justified it.
- A fact stored, in a bucket you name or the default one — `store_memory`
- A natural-language question answered from stored memories, with synthesis and a per-memory explanation — `query_memory`
- A bucket's memories listed newest first, with `limit` between 1 and 100 and a default of 20 — `list_memories`
- The buckets that exist — `list_buckets`
- One memory deleted by id, or a whole bucket cleared — `delete_memory`, `clear_memories`
- Model calls routed through your own provider, so retrieval carries no inference markup
A Lumetra account and an API key, sent as an `Authorization: Bearer` header to `https://mcp.lumetra.io/mcp/sse`. Claude.ai on the web and ChatGPT use OAuth instead and need no pasted key. There is no local installation. The free tier covers 10K stored memories and 50K retrievals per month with no credit card. One limit to plan around: the MCP `query_memory` tool takes a single bucket per call — multi-bucket fusion lives on the REST endpoint and in the SDKs.
One command plus a key — claude mcp add-json engram '{"type":"sse","url":"https://mcp.lumetra.io/mcp/sse","headers":{"Authorization":"Bearer <your-api-key>"}}', then supply credentials