Deployment

This page covers:

• the databases for Agentic Memory (PostgreSQL, Weaviate, ArangoDB, Redis) on Kubernetes;
• running the Learner and Incorporator services;
• running the test suites.

Download Image

---

Databases on Kubernetes

The manifests in agent-memories/k8s/ run all four databases in the agentic-memory namespace and expose them through NodePort services:

Database	Manifest	Image	Workload	Container port	NodePort	Persistent volume
PostgreSQL	postgres.yaml	postgres:16	StatefulSet	5432	30432	postgres-pvc (10 Gi)
ArangoDB	arango.yaml	arangodb:3.12	StatefulSet	8529	30529	arango-pvc (10 Gi)
Weaviate	weaviate.yaml	semitechnologies/weaviate:1.25.0	StatefulSet	8080 (HTTP), 50051 (gRPC)	30880, 30851	weaviate-pvc (10 Gi)
Redis	redis.yaml	redis:7	Deployment	6379	30637	none (data is not persisted)

Weaviate runs with anonymous access and no vectorizer module: Agentic Memory supplies its own embeddings. PostgreSQL, ArangoDB and Redis read their passwords from the postgres-secret, arango-secret and redis-secret Secrets defined in the manifests.

Steps

1. Create the host directories used by the hostPath volumes (on the node):

   bash mkdir -p /data/agentic-memory/{postgres,arango,weaviate}

2. Change the passwords in the Secrets in postgres.yaml, arango.yaml and redis.yaml.

3. Apply the manifests. Volumes go first, so the StatefulSets can bind their claims:

   bash kubectl create namespace agentic-memory # the manifests don't create the namespace kubectl apply -f k8s/pv-pvc.yaml kubectl apply -f k8s/postgres.yaml kubectl apply -f k8s/arango.yaml kubectl apply -f k8s/weaviate.yaml kubectl apply -f k8s/redis.yaml kubectl get pods -n agentic-memory -w

4. Point Agentic Memory at the node:

   ```python from agentic_memory import MemoryConfig, PostgresConfig, WeaviateConfig, ArangoConfig, RedisConfig

   NODE = "" config = MemoryConfig( postgres=PostgresConfig(host=NODE, port=30432, password=""), weaviate=WeaviateConfig(host=NODE, port=30880, grpc_port=30851), arango=ArangoConfig(url=f"http://{NODE}:30529", password=""), redis=RedisConfig(host=NODE, port=30637, password=""), ) ```

   The examples in agent-memories/examples/ read the same settings from environment variables: POSTGRES_HOST, POSTGRES_PORT, POSTGRES_PASSWORD, ARANGO_URL, ARANGO_PASSWORD, WEAVIATE_HOST, WEAVIATE_PORT, WEAVIATE_GRPC_PORT, REDIS_HOST, REDIS_PORT and REDIS_PASSWORD.

The PostgreSQL schema, Weaviate collections and ArangoDB collections are created automatically by the library on first connection.

The hostPath volumes suit single-node development clusters (minikube, kind, k3s). For production, use a StorageClass/CSI driver instead and do not expose the databases on public NodePorts.

---

Learner and Incorporator Services

Both are FastAPI apps. For a single machine, run them side by side with shared local storage:

export MG_TIDB_DSN="sqlite+aiosqlite:////srv/memorygrid/frame.db"
export MG_LESSONS_ROOT="/srv/memorygrid/lessons_repo"
export ANTHROPIC_API_KEY="..."              # optional; FakeLLMClient is used without it
export MG_LLM_MODEL="claude-sonnet-5-5"

(cd agent-memories/Learner && uvicorn memorygrid_learner.main:app --port 8000) &
(cd agent-memories/Incorporator && \
   MG_FEEDBACK_TIDB_DSN="sqlite+aiosqlite:////srv/memorygrid/feedback.db" \
   MG_DSL_MATCH_RULES_PATH="/srv/memorygrid/dsl_match_rules.json" \
   uvicorn memorygrid_incorporator.main:app --port 8001) &

For multiple machines or production, share real infrastructure:

Setting	Learner	Incorporator
MG_TIDB_DSN	TiDB/MySQL (mysql+aiomysql://…)	same
MG_VECTOR_BACKEND / MG_VECTOR_DSN	pgvector or milvus	same
MG_LESSONS_ROOT	Shared volume	same
MG_FEEDBACK_TIDB_DSN	—	Feedback DB

TiDB access needs a MySQL async driver (for example aiomysql), which is not in the requirements files. pgvector needs psycopg[binary] and pgvector; Milvus needs pymilvus.

Run one Learner instance, because the scheduler and watcher are in-process. Only the Learner runs the mining schedule and the draft watcher; the Incorporator does not.

---

Running the Tests

Package	Command	Needs services
Agentic Memory	cd agentic-memory && pytest	No (backends are mocked)
Agent FS	cd fs && pytest	No (needs git)
Learner	cd Learner && pytest -q	No (SQLite, in-memory vectors, fake LLM)
Incorporator	cd Incorporator && pytest -q	No (same, plus the sibling Learner/)