Persistence Strategies

Version: 0.1.0

ReasonKit Memory supports a dual-layer persistence strategy: Hot (Fast) and Cold (Archive).

1. Hot Storage (Vector Database)

Designed for sub-millisecond retrieval during active reasoning.

• Technology: Qdrant (primary), or pgvector (PostgreSQL).
• Data: Embeddings, metadata, recent ephemeral context.
• Retention: Configurable (e.g., last 30 days or active working set).

2. Cold Storage (Object/Relational)

Designed for durability, audit trails, and full reconstruction.

• Technology: SQLite (local), PostgreSQL (server), or S3-compatible Blob Storage.
• Data: Full raw text, original documents, complete conversation logs, snapshots of the vector state.
• Format: Parquet (for analytics) or JSONL (for portability).

Sync Strategy

1. Write Path:

   • Agent writes to MemoryInterface.
   • System writes to Cold Storage (WAL/Log) immediately for durability.
   • System asynchronously computes embeddings and updates Hot Storage.

2. Read Path:

   • Query hits Hot Storage (Vector Index).
   • If payload is missing/truncated in Hot, fetch full content from Cold Storage using ID.

Backup & Recovery

• Snapshotting: Qdrant snapshots are taken daily.
• PITR: PostgreSQL Point-in-Time Recovery is enabled for the Cold layer.
• Export: reasonkit-mem export --format jsonl allows dumping the entire memory state for migration.