Skip to content

novamem documentation

One memory across every AI agent you use. Hybrid keyword + vector + graph + recency + entity retrieval (5-signal engine, graph/entity defaulted to 0 in production calibration), per-user isolation with shareable sub-brains, MCP and HTTP transports, built-in dashboard. Integrates with 30+ AI agent hosts. Self-hostable on a laptop or as a multi-tenant brain for a whole company. pgvector as an alternative to Qdrant for the cold vector tier.

This is the long-form documentation. For the marketing landing page see novamem.github.io/novamem. For source see github.com/azrtydxb/novamem.

Where to start

What's here

SectionGoes deep on
InstallDocker Compose env reference, Kubernetes manifest walkthrough, manual Postgres + Qdrant setup
Connect agentsThe npx @azrtydxb/novamem-init CLI in detail; per-host (Claude Code, Desktop, ChatGPT, Cursor, Cline, Continue, Kilo, others); custom HTTP integration
DashboardSign-in & roles, every page tour, projects + sharing, tenant + user admin, API tokens
ArchitectureSystem shape, tiered storage, hybrid search internals, worthiness gate + dedup, decay maths + dream cycle, multi-tenancy
API referenceAuth flows, the data plane, admin & users, MCP tools, OpenAPI spec
OperationsSecurity model, hardening checklist, audit log, backup/restore, upgrades
ContributeLocal dev setup, project layout, testing, release flow, filing bugs

Three minutes overview

novamem is a single Fastify server that holds memory entries for AI agents and exposes two equivalent transports — JSON HTTP and the Model Context Protocol. Entries flow through five layers:

  • Warm tier — Postgres full-text search, low-latency hot path
  • Cold tier — Qdrant or pgvector vector embeddings, semantic recall over older entries
  • Relations — co-occurrence edges in Postgres (memory_relations), traversed by /v1/neighbors
  • Recency — exponential decay rank prior applied on search results
  • Entity bridge — exact identifier bridging via extracted identifiers from the query (currently at weight 0 in production)

Search fuses five signals — Postgres full-text keyword match, Qdrant/pgvector vector similarity, graph adjacency, recency rank prior, and entity bridge — with an optional cross-encoder rerank; adjacency between memories is served separately by /v1/neighbors over the memory_relations table.

The same code runs on a laptop (Docker Compose, single user) or as a multi-tenant deployment on Kubernetes, supporting 30+ AI agent hosts via the novamem-init CLI.