Persistence

By default, the supervisor agent uses in-memory storage for both conversation checkpoints and cross-thread memory. State is lost when the pod or container restarts.

This page explains how to enable durable persistence and, optionally, automatic fact extraction.

Concepts

Concept	What it stores	Default
Checkpoint	In-thread conversation state (messages, tool calls)	In-memory
Memory store	Cross-thread facts and summaries about each user	In-memory
Fact extraction	Automatically extracts facts from conversations and saves them to the store	Disabled

Supported backends: Redis Stack (recommended), PostgreSQL, MongoDB.

---

Option 1 — In-memory (default, no action required)

State is stored in-process. No external service is needed. Everything resets on restart.

This is the default. No configuration is required.

---

Option 2 — Redis persistence

Conversation checkpoints and cross-thread memories survive pod/container restarts.

Helm

Add the following to your values.yaml:

global:
  langgraphRedis:
    enabled: true          # deploy the bundled Redis Stack subchart

supervisor-agent:
  checkpointPersistence:
    type: redis
    redis:
      autoDiscoverService: langgraph-redis   # auto-builds redis://<release>-langgraph-redis:6379/0

  memoryPersistence:
    type: redis
    redis:
      autoDiscoverService: langgraph-redis

Or use the --persistence flag with setup-caipe.sh:

# Interactive — prompts "Enable Redis persistence?"
./setup-caipe.sh

# Non-interactive
./setup-caipe.sh --non-interactive --persistence

# With RAG stack
./setup-caipe.sh --non-interactive --rag --persistence

Docker Compose

Add to your .env:

LANGGRAPH_CHECKPOINT_TYPE=redis
LANGGRAPH_CHECKPOINT_REDIS_URL=redis://langgraph-redis:6379
LANGGRAPH_STORE_TYPE=redis
LANGGRAPH_STORE_REDIS_URL=redis://langgraph-redis:6379

Then start with the langgraph-redis profile:

docker compose -f docker-compose.dev.yaml --profile langgraph-redis up

Or use the switch_backend.sh helper:

./skills/persistence/switch_backend.sh redis

---

Option 3 — Redis persistence + fact extraction

Same as Option 2, plus the supervisor automatically extracts facts from every conversation turn and stores them for future cross-thread recall.

Note: Fact extraction makes one additional LLM call per conversation turn. Ensure your LLM endpoint has sufficient capacity.

Helm

global:
  langgraphRedis:
    enabled: true

supervisor-agent:
  checkpointPersistence:
    type: redis
    redis:
      autoDiscoverService: langgraph-redis

  memoryPersistence:
    type: redis
    redis:
      autoDiscoverService: langgraph-redis
    enableFactExtraction: true   # extract facts from every conversation turn
    maxMemories: 50              # max facts recalled per cross-thread query (default: 50)
    maxSummaries: 10             # max summaries recalled per cross-thread query (default: 10)

Docker Compose

Add to your .env:

LANGGRAPH_CHECKPOINT_TYPE=redis
LANGGRAPH_CHECKPOINT_REDIS_URL=redis://langgraph-redis:6379
LANGGRAPH_STORE_TYPE=redis
LANGGRAPH_STORE_REDIS_URL=redis://langgraph-redis:6379
ENABLE_FACT_EXTRACTION=true

Then start with the langgraph-redis profile:

docker compose -f docker-compose.dev.yaml --profile langgraph-redis up

---

Bring your own Redis (BYO)

If you already have a Redis Stack instance (e.g. AWS ElastiCache for Redis, an existing in-cluster Redis), skip global.langgraphRedis.enabled and point directly at your instance.

Requirement: Redis Stack with RedisJSON and RediSearch modules is required. Plain Redis without these modules will not work.

Helm — direct URL

supervisor-agent:
  checkpointPersistence:
    type: redis
    redis:
      url: "redis://my-redis.example.com:6379/0"

  memoryPersistence:
    type: redis
    redis:
      url: "redis://my-redis.example.com:6379/0"

Helm — from a Kubernetes Secret (recommended for credentials)

supervisor-agent:
  checkpointPersistence:
    type: redis
    redis:
      existingSecret:
        name: my-redis-secret
        key: REDIS_URL

  memoryPersistence:
    type: redis
    redis:
      existingSecret:
        name: my-redis-secret
        key: REDIS_URL

---

Reusing the RAG Redis

The RAG stack includes its own Redis (rag-redis). You can reuse it for persistence by setting autoDiscoverService: rag-redis and using a separate DB index to avoid collision with RAG data.

Requirement: The RAG Redis image must be redis:8.0+ (includes RedisJSON and RediSearch natively). The default redis:7.2-alpine image does not include these modules.

supervisor-agent:
  checkpointPersistence:
    type: redis
    redis:
      autoDiscoverService: rag-redis
      dbIndex: 1    # db 0 is used by RAG data

  memoryPersistence:
    type: redis
    redis:
      autoDiscoverService: rag-redis
      dbIndex: 1

---

Full configuration reference

supervisor-agent:
  memoryPersistence:
    type: "memory"            # memory | redis | postgres | mongodb
    redis:
      url: ""                 # direct URL
      existingSecret:         # pull URL from a Kubernetes Secret
        name: ""
        key: ""
      autoDiscoverService: "" # "langgraph-redis" or "rag-redis"
      dbIndex: 0
      keyPrefix: ""           # namespace prefix for shared Redis
    postgres:
      dsn: ""
      existingSecret:
        name: ""
        key: ""
    mongodb:
      uri: ""
      existingSecret:
        name: ""
        key: ""
    embeddings:
      provider: ""            # openai | azure-openai | litellm
      model: ""               # e.g. text-embedding-3-small
      dims: ""                # auto-detected for known models
    ttlMinutes: 10080         # 7 days; 0 = no expiry
    enableFactExtraction: false
    maxMemories: 50
    maxSummaries: 10

  checkpointPersistence:
    type: "memory"            # memory | redis | postgres | mongodb
    redis:
      url: ""
      existingSecret:
        name: ""
        key: ""
      autoDiscoverService: ""
      dbIndex: 0
    postgres:
      dsn: ""
      existingSecret:
        name: ""
        key: ""
    mongodb:
      uri: ""
      existingSecret:
        name: ""
        key: ""
    ttlMinutes: 0             # 0 = no expiry

global:
  langgraphRedis:
    enabled: false            # set true to deploy the bundled Redis Stack subchart