Memorylayer keeps hosted concerns separate from Engram: identity, workspaces, API keys, audit trails, connection kits, and a small HTTP bridge. The memory runtime is Postgres-backed Engram with explicit embedding and reranker models.
Codex, Claude, custom runners, CI jobs, and scripts call a workspace-scoped surface.
GitHub login, workspace membership, hashed keys, audit trails, and usage events.
Every workspace maps to its own Engram schema and operational history.
Retrieval, graph, handoff, curation, and continuity stay inside the real engine.
The live service is a narrow hosted control plane around the Engram runtime. These are the concrete parts currently wired in this repo.
FastAPI app served by Uvicorn on port 8090 inside Docker, exposed through https://memorylayer.run.
Postgres stores service metadata, users, workspaces, keys, usage events, audit events, ingest runs, and each workspace's Engram schema.
engram-memory-system 0.5.2 provides the store, retrieval, graph, curation, session handoff, and MCP-compatible tool layer.
These are the memory models exposed by Engram's current config defaults, with the hosted service overriding storage to Postgres.
BAAI/bge-small-en-v1.5 creates dense vectors for semantic memory search.384 dimensions per embedding vector.cross-encoder/ms-marco-MiniLM-L-6-v2 is the cross-encoder reranker model when reranking is available.embedding_backend = auto; the hosted runtime uses CPU PyTorch container.sqlite, but Memorylayer sets storage_backend = postgres and injects a workspace schema DSN.The hosted service does not pretend to be the memory system. It gives the memory system accounts, keys, URLs, and operational controls.
GitHub OAuth, sessions, members, invites, and dashboard routes.
API keys, audit events, usage events, ingestion runs, exports, and connection kits.
Engram recall, remember, graph, curation, summaries, handoffs, and health checks.
A workspace API call has a small number of predictable checkpoints.
Simple by design: one Python app, one Postgres database, one long-lived memory process model.
Public pages, workspace dashboard, OAuth, and JSON endpoints.
App metadata plus per-workspace Engram schemas with postgres search_path is scoped per workspace schema.
Repeatable VPS deployment using python:3.12-slim and a narrow exposed port.
2 starter skills, 5 playbooks, and 6 SDK snippets.
The app keeps the operational shape explicit so the memory process stays predictable.
16 workspace runtimes are kept warm in-process.1800 seconds.2000000 bytes before routing.240 requests per minute per client bucket.12 requests per minute per client bucket.