One memory across every AI agent you use.

LLMs forget. RAG isolates. Each agent reinvents.

A 200K-token context isn't memory — it's a goldfish bowl that resets every session. The usual workaround, "stuff a vector DB behind it," misses everything you actually want from memory. And the moment you have more than one agent, every one of them ends up with its own private notes.

Different agents. Same memory.

The agent landscape moves fast. Whatever you and your team use this quarter — and whatever you switch to next quarter — they all read and write the same novamem store. No re-onboarding. No silos. One canonical place where everything goes.

The mental model — and the seven tools.

Every entry belongs to a single user. An entry can additionally belong to a project — a sub-brain that's shareable. User-global entries are private; project entries are visible to every member of that project. Every search runs three signals in parallel and fuses them into one ranked list.

Full usage guide → covers worthiness gates, decay maths, dream cycle, namespaces, and weight tuning.

Batteries included. No vendor SaaS.

One docker compose up -d brings the whole stack online. No external services. No proprietary embedding API. No hidden per-token charges.

Three retrieval signals, fused. One coherent answer.

Every search runs keyword (FTS), vector (cosine), and graph (neighbour traversal) in parallel. Results fuse via min-max-normalised weighted scoring with sensible defaults you can override per call.

What each tier does — concretely.

A simple example. Imagine you've been remembering project-notes for weeks. Today you ask your agent: "How did we end up choosing Postgres for the main store?" Here's what each tier contributes — and why fusing all three beats any one alone.

All three signals run in parallel and fuse into one ranked list. The warm hit gets you the literal ADR. The cold hit pulls in the prior reasoning even though it never said "Postgres". The graph hit ties the decision to its supporting cost analysis. Your agent answers with all of it — not just the one tier that happened to match.

Stand it up. Three paths.

Pick whichever fits your environment. All three lead to the same server image, the same dashboard, the same MCP surface.

# clone, set 3 secrets, up
git clone https://github.com/azrtydxb/novamem.git
cd novamem && cp .env.example .env

echo "POSTGRES_PASSWORD=$(openssl rand -base64 24)" >> .env
echo "NOVAMEM_BOOTSTRAP_ADMIN_PASSWORD=$(openssl rand -base64 24)" >> .env
echo "NOVAMEM_COOKIE_SECRET=$(openssl rand -hex 32)" >> .env

docker compose up -d
# http://localhost:7778/admin

# bring your own Postgres + Qdrant + FalkorDB
# prereqs: Node 20+, pnpm 9+
git clone https://github.com/azrtydxb/novamem.git
cd novamem
pnpm install && pnpm build

cp .env.example .env
# point at your existing datastores
pnpm --filter @azrtydxb/novamem-server start

# multi-arch image on ghcr.io
# manifests in deploy/k8s/
git clone https://github.com/azrtydxb/novamem.git
cd novamem/deploy/k8s

# edit secrets.yaml + ingress.yaml host
# then:
kubectl apply -k .
kubectl -n novamem rollout status deploy/novamem

Wire every AI host on your machine. One command.

@azrtydxb/novamem-init detects 30+ supported hosts, asks for your server URL + dashboard credentials, mints a fresh bearer, and writes the config each host expects. Idempotent — won't clobber existing entries.

npx -y @azrtydxb/novamem-init

# detects: Claude Code · Claude Desktop · ChatGPT (via HTTP)
#          Cursor · OpenCode · Codex CLI · Cline · Continue
#          Kilo Code · RooCode · Gemini CLI · Copilot · Windsurf
#          Factory · Amazon Q · & ~16 skill-only hosts

Prefer manual setup? Per-host walkthroughs: Claude Code · Claude Desktop · Cursor · Kilo Code · Other hosts & skills

Same product, three deploys.

Same code. Same MCP surface. Same dashboard. The only thing that changes between a personal laptop and a 5,000-engineer company is how you stand it up. Multi-tenancy and project-based sub-brains are first-class from day one.

Specs & references.

Everything you need to operate, integrate, or audit. The OpenAPI spec is generated from the same Zod schemas the server uses at runtime — so it's always accurate.

The decisions, on the table.

novamem is opinionated about a few things — and the rest is yours.

Give your agents a memory.

Self-host in a minute. Wire every AI host on your machine in one npx command.