What Is HyperMemory?

An AI backend that stores memories — and actually knows how to find them again.

The Problem It Solves

AI assistants forget everything between conversations. HyperMemory is the persistent memory layer that gives them a real past.

When You Store a Memory

Imagine you send this to the API endpoint POST /mcp/memory/store:

Here is what happens next — click through to trace the journey:

Seven Generations of Getting Smarter

HyperMemory didn't start this complex. It was built layer by layer — each version adding a new superpower on top of the last.

The Database Trio

Three completely different databases, each solving a problem the others can't.

PostgreSQL — The Source of Truth

Everything real lives here. Memories, extracted facts, entities, relationships, tiers. If it's deleted from PostgreSQL, it's gone.

class Memory(Base):
    __tablename__ = "memories"
    id = Column(Integer, primary_key=True)
    content = Column(Text, nullable=False)
    summary = Column(Text)
    user_id = Column(String, index=True)
    importance = Column(Float, default=0.5)
    timestamp = Column(DateTime)
          

Qdrant — The Concept Library

Qdrant stores memories as vectors — numerical fingerprints of meaning. When you search, it finds the memories whose meaning is closest to your query.

Redis — The Quick-Reference Board

Redis holds recent search results in memory so they can be returned instantly, without re-running expensive searches. Like a whiteboard next to the filing cabinet — frequently needed answers written in marker, cleared at the end of the day.

Watch how these three databases collaborate on a typical query:

Check Your Understanding

The Extraction Factory

How raw text becomes structured knowledge — automatically, in the background, without blocking you.

The Background Worker

Extraction doesn't block your API call. An async worker runs in the background, polling for unprocessed memories every 30 seconds.

class ExtractionWorker:
    def __init__(self,
        batch_size: int = 10,
        poll_interval: int = 30,
        graph_extraction_enabled: bool = True,
        event_segmentation_enabled: bool = True,
    ):
          

Three Extraction Pipelines

The ExtractionWorker runs three separate pipelines on each memory:

Inside the Extraction Pipeline

Here is what happens inside the worker when it picks up a new memory:

Check Your Understanding

The Search Brain

Five strategies. One question. The right answer — even when the words don't match.

A Query Travels Through Seven Stages

Before any database is touched, your query goes through analysis, expansion, and classification. Here is the full pipeline:

Two Types of "Similar"

Keyword search and semantic search find different things. HyperMemory runs both and blends the results.

vector_engine = RetrievalEngine(
    postgres_adapter=db_adapter,
    vector_adapter=vector_adapter,
    synthesis_model=settings.EXTRACTION_MODEL,
    synthesis_base_url=settings.EXTRACTION_BASE_URL,
    synthesis_api_key=settings.EXTRACTION_API_KEY,
)
          

PRISM — The Research Robot

For complex questions that need multiple memory pieces to answer, v7 adds PRISM: a 3-agent loop that keeps refining until it has exactly the right evidence.

Check Your Understanding

Memory That Lives and Dies

Not all memories are equal. HyperMemory knows which ones to keep, which to fade, and which to bury.

The Heat Score — A Memory's Life Force

Every memory has a heat score between 0.0 and 1.0. High heat = hot memory, actively used, should be kept close. Low heat = cold memory, forgotten, getting moved to the archive.

visit_score = min(1.0, math.log10(visit_count + 1) / 3.0)
recency_score = math.exp(-hours_since / 168)
heat = (
    0.4 * visit_score
    + 0.3 * recency_score
    + 0.3 * importance_score
)
          

The Memory Tier Ladder

Memories are constantly being promoted and demoted based on their heat score. A background worker runs every hour and reorganizes them.

The Hourly Gardening Round

Every hour, the DecayScheduler wakes up and runs its batch. Watch what happens:

Check Your Understanding