langgraph-store-pixeltable

LangGraph BaseStore backend for Pixeltable — persistent, versioned, multimodal agent memory.

Installation

pip install langgraph-store-pixeltable

Quick Start

from langgraph.store.pixeltable import PixeltableStore

store = PixeltableStore(table_name="agent_memory.items")
store.setup()

# Store a memory
store.put(("users", "alice"), "prefs", {"color": "blue", "language": "Python"})

# Retrieve it
item = store.get(("users", "alice"), "prefs")
print(item.value)  # {"color": "blue", "language": "Python"}

# List namespaces
namespaces = store.list_namespaces(prefix=("users",))

Filtered Search

The filter parameter maps to Pixeltable's .where() clause — predicates are evaluated server-side, not in Python:

store.put(("team", "eng"), "alice", {"role": "senior", "level": 5, "lang": "Python"})
store.put(("team", "eng"), "bob", {"role": "junior", "level": 2, "lang": "Go"})
store.put(("team", "eng"), "carol", {"role": "senior", "level": 4, "lang": "Python"})

# Exact match (shorthand)
results = store.search(("team",), filter={"role": "senior"})

# Comparison operators
results = store.search(("team",), filter={"level": {"$gt": 3}})
results = store.search(("team",), filter={"role": {"$ne": "junior"}})

# Multiple conditions (AND)
results = store.search(("team", "eng"), filter={"role": "senior", "lang": "Python"})

Supported operators: $eq, $ne, $gt, $gte, $lt, $lte.

Semantic Search

from sentence_transformers import SentenceTransformer

model = SentenceTransformer("all-MiniLM-L6-v2")

store = PixeltableStore(
    table_name="agent_memory.semantic",
    index={
        "dims": 384,
        "embed": lambda texts: model.encode(texts).tolist(),
        "fields": ["text"],
    },
)
store.setup()

store.put(("docs",), "d1", {"text": "Pixeltable handles multimodal data", "domain": "infra"})
store.put(("docs",), "d2", {"text": "LangGraph builds stateful agents", "domain": "framework"})
store.put(("docs",), "d3", {"text": "React is a frontend library", "domain": "frontend"})

# Semantic search
results = store.search(("docs",), query="multimodal AI pipelines", limit=2)
for r in results:
    print(f"{r.key}: {r.value['text']} (score={r.score:.3f})")

# Semantic search + filter (narrowed by domain)
results = store.search(
    ("docs",), query="data infrastructure", limit=3, filter={"domain": "infra"},
)

With LangGraph Agents

from langgraph.prebuilt import create_react_agent
from langgraph.store.pixeltable import PixeltableStore

store = PixeltableStore(table_name="agent_memory.items")
store.setup()

agent = create_react_agent(model, tools=tools, store=store)

Access the Underlying Pixeltable Table

The .table property gives direct access to the Pixeltable table for operations beyond the Store interface — computed columns, lineage, version history, and arbitrary predicates:

import pixeltable as pxt

t = store.table

# Inspect all data
t.select(t.key, t.value, t.created_at).collect()

# Add a computed column — auto-backfills all existing rows
t.add_computed_column(word_count=my_word_counter(t.value['text'].astype(pxt.String)))

# Add a classification UDF — runs on every new insert
t.add_computed_column(auto_domain=classify_domain(t.value['text'].astype(pxt.String)))

# WHERE on computed columns
results = t.where(t.auto_domain == 'infra').select(t.key, t.value).collect()

# Compound WHERE on multiple computed columns
results = (
    t.where((t.word_count > 5) & (t.auto_domain == 'infra'))
    .select(t.key, t.value, t.word_count, t.auto_domain)
    .collect()
)

# New inserts via the store auto-compute all lineage columns
store.put(("docs",), "d4", {"text": "Kubernetes orchestrates containers"})
# word_count and auto_domain are already computed for d4

Why Pixeltable for LangGraph Memory?

• Metadata filtering via .where(): Filter on value fields with $eq/$ne/$gt/$gte/$lt/$lte, evaluated server-side
• Computed column lineage: Add derived columns that auto-backfill and auto-compute on new inserts
• Persistent and versioned: Data survives restarts; every update is tracked with full version history
• Incremental: Only new/changed rows get re-embedded
• Multimodal native: Images, video, audio, and documents alongside text via .table
• Any embedding model: Works with sentence-transformers, OpenAI, or any callable
• No external services: Embedded PostgreSQL, no Docker required

Feature	PostgresStore	PixeltableStore
Persistence	Yes	Yes + versioned
Vector search	pgvector only	Any embedding model
Multimodal values	JSON only	Image, Video, Audio via .table
Incremental embedding	Manual	Automatic via computed columns
History	No	Full version history per row
Computed columns	No	Arbitrary UDFs with lineage
External service	Requires PostgreSQL + pgvector	Embedded, no Docker

Links

• Pixeltable Docs
• LangGraph Integration Docs
• GitHub
• Discord