> ## Documentation Index > Fetch the complete documentation index at: https://praison.ai/docs/llms.txt > Use this file to discover all available pages before exploring further. # Memory Storage > Database backends for agent memory and state persistence # Memory Storage PraisonAI supports 22+ database backends for memory persistence. Choose the right one for your use case. ## Storage Categories ### Conversation Stores Store conversation history and chat context. Best for session continuity. Production-ready relational DB Popular relational DB Zero-config local storage Simple file-based storage Managed Postgres Serverless Postgres ### State Stores Store agent state, checkpoints, and runtime data. Best for complex workflows. High-speed caching Document store AWS managed Google Cloud Serverless Redis Decay-weighted vector recall ## Quick Setup ```python theme={"theme":{"light":"vitesse-light","dark":"vitesse-dark"}} from praisonaiagents import Agent # SQLite (zero-config, default) agent = Agent(memory=True) # SQLite via preset agent = Agent(memory="sqlite") # MongoDB via URL scheme agent = Agent( memory={ "db": "mongodb://localhost:27017/mydb", "session_id": "user-123" } ) ``` `memory=` presets and URL schemes accept only shipped memory adapters. Presets: `chroma`, `dakera`, `file`, `in_memory`, `mem0`, `mongodb`, `sqlite`. URL schemes: `sqlite://`, `mongodb://`, `mongodb+srv://`. Values like `memory="redis"` or `memory={"backend": "redis"}` raise `ValueError`. For Redis/Postgres *state* persistence, use the storage backends below or `db(state_url="redis://…")`. ## Choosing a Database | Use Case | Recommended | Why | | ------------------------------------- | ------------ | --------------------------------------------- | | Local development | SQLite/JSON | Zero config | | Production web app | PostgreSQL | Reliable, scalable | | High-speed caching | Redis | Sub-ms latency | | Serverless | Neon/Upstash | No server management | | AWS infrastructure | DynamoDB | Native integration | | Decay-weighted recall across sessions | Dakera | Importance-scored, time-decayed memory server | For self-hosted decay-weighted vector memory with importance scoring and tiered recall, see [Dakera Memory](/docs/features/dakera-memory). ## Storage Backends For training data, sessions, and general persistence, PraisonAI provides pluggable storage backends: | Backend | Dependencies | Best For | | ----------------- | ------------ | ------------------------------- | | **FileBackend** | None | Development, debugging | | **SQLiteBackend** | None | Production, concurrent access | | **RedisBackend** | `redis` | High-speed caching, distributed | ```python theme={"theme":{"light":"vitesse-light","dark":"vitesse-dark"}} from praisonaiagents.storage import FileBackend, SQLiteBackend, RedisBackend # File-based (default) file_backend = FileBackend(storage_dir="~/.praisonai/data") # SQLite (recommended for production) sqlite_backend = SQLiteBackend(db_path="~/.praisonai/data.db") # Redis (for distributed systems) redis_backend = RedisBackend(url="redis://localhost:6379", prefix="praison:") ``` FileMemory sanitises `user_id` before using it as a directory name — values containing path separators, `..`, or other unsafe characters fall back to `"default"`. This prevents accidental writes outside the configured memory directory. ### Durability `FileMemory` writes are atomic. Every `_write_json` writes into a same-directory temp file, `fsync`s it, then `os.replace`s it into place — a crash mid-write (SIGKILL, power loss, OOM) can never leave the memory file empty or half-written. Same guarantee on POSIX and Windows. * `flock`-based mutual exclusion still applies on Unix so concurrent writers see a consistent view (Windows has no `flock`; atomicity of `os.replace` still holds). * Temp files use a `..…tmp` prefix inside the memory directory. A crash between temp creation and rename leaves the temp file behind; it is safe to delete manually and is otherwise harmless. See [Storage Backends](/docs/storage/backends) for detailed usage with all components. ## Related Complete database reference Continue conversations Pluggable storage backends Self-hosted decay-weighted vector memory