LangGraph memory mechanisms (mechanism × persistence)

langgraph-memory-mechanisms.md

LangGraph memory mechanisms (mechanism × persistence)

1. Execution / short-term memory (checkpointers)

Mechanism	Purpose	In-memory / local	DB-persisted	Notes
MemorySaver	Persist graph state per step (thread continuity)	✅ Yes	❌ No	Notebook/dev only; lost on process restart
PostgresSaver	Durable execution state	❌ No	✅ Postgres	Required for resumable workflows, HITL, failures
(others)	(Future / custom savers)	—	—	Checkpointer interface is pluggable

What it stores

• Full graph state
• Current node
• Messages / flags / intermediate state

Key idea

Checkpointers persist execution, not knowledge.

---

2. Long-term / cross-thread memory (Store API)

Mechanism	Purpose	In-memory / local	DB-persisted	Notes
InMemoryStore	Cross-thread memory items	✅ Yes	❌ No	Process-local KV store
PostgresStore	Durable memory items	❌ No	✅ Postgres	Production default
(custom store)	S3, Redis, etc.	Depends	Depends	Implement BaseStore

What it stores

• Arbitrary key/value objects
• Namespaced by tuples (e.g. (user_id, "profile"))
• Optional embedding indexes

Key idea

Stores persist agent memory, not execution flow.

---

3. Semantic memory (Store + embeddings)

Mechanism	Purpose	In-memory / local	DB-persisted	Notes
InMemoryStore + embeddings	Semantic recall	✅ Yes	❌ No	Embeddings live in RAM
PostgresStore + embeddings	Durable semantic memory	❌ No	✅ Postgres	Requires vector support
External vector DB (pattern)	Large-scale semantic memory	❌ No	✅ External	LangGraph does not mandate one

What it stores

• Text or structured blobs
• Vector embeddings
• Retrieved via store.search()

Key idea

Semantic memory is a capability of the store, not a separate primitive.

---

4. Episodic memory (pattern, not a primitive)

Mechanism	Purpose	In-memory / local	DB-persisted	Notes
Store-backed episodes	Recall past interactions	✅ Yes (InMemoryStore)	✅ Yes (PostgresStore)	Pattern implemented on top of Store
Few-shot injection	Condition LLM behavior	N/A	N/A	Prompt-time use only

What it stores

• Past interactions
• Outcomes / feedback
• Free-form or structured episode records

Key idea

Episodic memory is modeled data + retrieval, not a LangGraph primitive.

---

5. Procedural memory (pattern, not a primitive)

Mechanism	Purpose	In-memory / local	DB-persisted	Notes
Store-backed instructions	Agent behavior policy	✅ Yes	✅ Yes	Often versioned text blobs
Reflection node	Self-modification	N/A	N/A	Writes updated policy to Store

What it stores

• Instructions
• Policies
• Strategy hints

Key idea

Procedural memory is policy state, externalized via Store.

---

Cross-cutting summary (the important part)

Where memory actually lives

Memory category	Lives inside graph?	Lives outside graph?
Graph definition	❌	❌ (pure code)
Execution state	❌	✅ (checkpointer)
Agent memory	❌	✅ (store)
Domain knowledge	❌	❌ (should live elsewhere, e.g. Graphiti)

LangGraph graphs are stateless. All memory is external.

---

Minimal production guidance (rule of thumb)

Use case	Required components
Chat continuity	PostgresSaver
User profiles / preferences	PostgresStore
Semantic RAG	PostgresStore or external vector DB
Learning from experience	PostgresStore
World model / facts	Graphiti (separate system)