cellin

Cellin builds long-lived multimodal memory, dreams over it to consolidate ideas, and retrieves context with transparent weighted ranking.

Install

From source today:

git clone https://github.com/ben-ranford/cellin.git
cd cellin
make bootstrap

Install from PyPI:

python3 -m pip install cellin

Run as MCP server

Cellin can run as an MCP stdio server for clients such as Claude Desktop and Cursor:

python3 -m pip install "cellin[mcp]"
cellin mcp serve --check
cellin mcp serve

The Docker image starts the same stdio server by default:

make docker-build
docker run -i --rm -v cellin-data:/data cellin-mcp:latest

Environment variables select packaged storage presets:

Variable	Default	Purpose
CELLIN_BACKEND	sqlite	sqlite, postgres, neo4j, or in_memory
CELLIN_DATA_DIR	/data	SQLite subject database and subject index directory
CELLIN_CONNECTION_STRING	unset	PostgreSQL DSN or Neo4j bolt URL for remote presets

For PostgreSQL:

docker run -i --rm \
  -e CELLIN_BACKEND=postgres \
  -e CELLIN_CONNECTION_STRING="$CELLIN_POSTGRES_DSN" \
  cellin-mcp:latest

For Neo4j:

docker run -i --rm \
  -v cellin-data:/data \
  -e CELLIN_BACKEND=neo4j \
  -e CELLIN_CONNECTION_STRING="$CELLIN_NEO4J_URI" \
  cellin-mcp:latest

docker-compose.yml includes a single-container SQLite service and profiled PostgreSQL and Neo4j variants:

docker compose up --build cellin-sqlite
CELLIN_POSTGRES_PASSWORD="$CELLIN_POSTGRES_PASSWORD" \
CELLIN_POSTGRES_DSN="$CELLIN_POSTGRES_DSN" \
  docker compose --profile postgres up --build cellin-postgres
CELLIN_NEO4J_AUTH="$CELLIN_NEO4J_AUTH" \
CELLIN_NEO4J_URI="$CELLIN_NEO4J_URI" \
  docker compose --profile neo4j up --build cellin-neo4j

Claude Desktop configuration:

{
  "mcpServers": {
    "cellin": {
      "command": "docker",
      "args": ["run", "-i", "--rm", "-v", "cellin-data:/data", "cellin-mcp:latest"]
    }
  }
}

Quickstart

From the repository root:

WORKSPACE=.cellin-workspace
python3 -m uv run cellin init --workspace "$WORKSPACE"
python3 -m uv run cellin storage list --role memory
python3 -m uv run cellin storage init --config "$WORKSPACE/cellin.json" --dry-run
python3 -m uv run cellin ingest --config "$WORKSPACE/cellin.json" --input examples/starter/seed_envelopes.json
python3 -m uv run cellin retrieve --config "$WORKSPACE/cellin.json" --query "memory graph retrieval" --top-k 2
python3 -m uv run cellin dream --config "$WORKSPACE/cellin.json" --strategy abstraction
python3 -m uv run cellin eval run --suite smoke --config "$WORKSPACE/cellin.json" --output "$WORKSPACE/smoke.json"
python3 -m uv run cellin trace inspect --config "$WORKSPACE/cellin.json" --limit 5

See examples/starter/README.md for the same flow in a shorter checklist form.

Workspace config now supports role-specific storage backends:

{
  "runtime_id": "cellin-cli",
  "trace_path": "traces.jsonl",
  "profile_name": "balanced",
  "storage": {
    "memory": { "backend": "in_memory" },
    "graph": { "backend": "in_memory" },
    "vector": { "backend": "in_memory_vector_index" },
    "representation": { "backend": "in_memory_vector_index" }
  }
}

cellin init now writes this in-memory-first preset by default.

Legacy workspaces that only define database_path continue to work and are migrated to this shape behind the scenes.

Use cellin storage list to inspect the built-in provider registry and any third-party providers registered through the cellin.storage entry-point group. Run cellin storage init --config <workspace>/cellin.json before first use when you want an explicit, idempotent setup step for durable backends.

For an explicit SQLite preset, set:

{
  "memory": { "backend": "sqlite", "database_path": "cellin.sqlite" },
  "graph": { "backend": "sqlite", "database_path": "cellin.sqlite" }
}

For additional SQL-backed presets, install optional dependencies as needed:

python3 -m pip install cellin[duckdb]
python3 -m pip install cellin[postgresql]
python3 -m pip install cellin[mysql]
python3 -m pip install cellin[sql-backends]
python3 -m pip install cellin[storage-backends]

Use duckdb to point both roles at a local DB file:

{
  "memory": { "backend": "duckdb", "database_path": "cellin.duckdb" },
  "graph": { "backend": "duckdb", "database_path": "cellin.duckdb" }
}

Use postgresql and mysql with connection strings:

{
  "memory": { "backend": "postgresql", "database_path": "postgresql://db.example.invalid:5432/db" },
  "graph": { "backend": "postgresql", "database_path": "postgresql://db.example.invalid:5432/db" }
}

{
  "memory": { "backend": "mysql", "database_path": "mysql://db.example.invalid:3306/db" },
  "graph": { "backend": "mysql", "database_path": "mysql://db.example.invalid:3306/db" }
}

For document and cache-oriented presets, install optional dependencies as needed:

python3 -m pip install cellin[mongodb]
python3 -m pip install cellin[redis]
python3 -m pip install cellin[document-cache-backends]

Use mongodb when you want durable document storage for both memories and edges:

{
  "memory": { "backend": "mongodb", "database_path": "mongodb://db.example.invalid:27017/cellin" },
  "graph": { "backend": "mongodb", "database_path": "mongodb://db.example.invalid:27017/cellin" }
}

Use redis for low-latency cache-oriented deployments where operators control TTL or eviction:

{
  "memory": { "backend": "redis", "database_path": "redis://host:6379/0" },
  "graph": { "backend": "redis", "database_path": "redis://host:6379/0" }
}

MongoDB stores whole memory and edge documents and preserves archived entries as tombstones in the document payloads. Redis stores JSON payloads per key and also preserves archived entries as tombstones, filtering them from neighbor and edge listing reads rather than hard-deleting them.

For graph-native backends, install the optional dependencies you need:

python3 -m pip install cellin[neo4j]
python3 -m pip install cellin[memgraph]
python3 -m pip install cellin[arangodb]
python3 -m pip install cellin[graph-backends]

Use neo4j, memgraph, or arangodb for the graph role while keeping memory storage separate if you prefer:

{
  "memory": { "backend": "sqlite", "database_path": "cellin.sqlite" },
  "graph": { "backend": "neo4j", "database_path": "bolt://graph.example.invalid:7687" }
}

Graph-native stores persist edge relationships plus graph-local memory payload snapshots. When GraphStore.get_memory() is asked for a node that only exists as a placeholder created during edge upserts, it returns None and retrieval falls back to the configured memory store as the source of truth. That keeps mixed deployments working without caller changes while making the graph-local snapshot behavior explicit.

For vector-backed retrieval stores, install the optional dependencies you need:

python3 -m pip install cellin[pgvector]
python3 -m pip install cellin[pinecone]
python3 -m pip install cellin[qdrant]
python3 -m pip install cellin[weaviate]
python3 -m pip install cellin[milvus]
python3 -m pip install cellin[redis-vector]
python3 -m pip install cellin[vector-backends]

Use cellin storage init --config "$WORKSPACE/cellin.json" after switching a workspace onto a durable backend family so local files, schemas, or remote collection bootstrap happen explicitly before ingest or retrieval.

Primary surfaces

• CLI: cellin init, storage list, storage init, ingest, retrieve, dream, plugin list, eval run, trace inspect
• Ingestion: cellin.ingest.ArtifactEnvelope and cellin.ingest.CanonicalIngestor
• Retrieval: cellin.retrieval.WeightedRetriever, cellin.retrieval.RetrievalCandidateGenerator, and cellin.ranking.WeightedRanker
• Dreaming: cellin.dreaming.DreamRunner plus the built-in deduplication, abstraction, and contradiction-repair strategies
• Evals: cellin.evals.run_evaluation_suite and cellin.evals.run_smoke_eval
• Extensibility: cellin.runtime.PluginRegistry and the contracts exported from cellin.core