Memory Operations

CortexDB provides five core operations for working with memory. Each operation is designed around how AI systems naturally interact with information.

remember

Store a new episode in CortexDB.

remember stores an episode durably. Once you receive an acknowledgment, the data is safe. Additional processing (entity discovery, indexing) happens asynchronously in the background.

result = client.remember(
    content="We should switch from REST to gRPC for the internal APIs.",
    tenant_id="my-app",
    metadata={"source": "slack", "author": "alice"},
)

print(result.event_id)   # Unique event ID

Key characteristics:

• Fast writes -- no LLM on the write path
• Content is preserved losslessly -- never rewritten
• Idempotent when the same episode ID is provided

Batch Ingest

For high-throughput ingestion, use ingest_episode with structured episodes:

from cortexdb import IngestEpisodeRequest

episode = IngestEpisodeRequest.from_dict({
    "content": "We should switch from REST to gRPC.",
    "type": "decision",
    "source": "slack",
    "actor": "alice",
})
result = client.ingest_episode(episode)

recall

Retrieve relevant memories using natural language.

recall is the primary read operation. It finds the most relevant context matching your query.

results = client.recall(
    query="What API protocol are we using internally?",
    tenant_id="my-app",
)

print(results.context)       # Synthesized context
print(results.confidence)    # Confidence score

forget

Remove memories from recall results.

forget removes information from all read paths. The deletion is auditable and compliant.

result = client.forget(
    query="old migration notes",
    reason="Decision reversed",
    tenant_id="my-app",
)

print(f"Forgotten {result.forgotten_entities} entities, {result.forgotten_edges} edges")

Key characteristics:

• Immediately excluded from recall and search results
• Auditable -- every forget operation is tracked

search

Search across memories and episodes with filters.

search provides filtered access to episodes when you need fine-grained control.

# Search with filters
results = client.search(
    query="authentication issues",
    source="slack",
    episode_type="message",
    time_range="7d",
    tenant_id="my-app",
)

# Search for exact terms
results = client.search(
    query="CORTEX-789",
    tenant_id="my-app",
)

link

Manually create relationships between entities.

While CortexDB automatically discovers entity relationships, link lets you explicitly declare connections that represent domain-specific knowledge.

result = client.link(
    source_entity_id="ent_payments_service",
    target_entity_id="ent_cockroachdb",
    relationship="uses",
    fact="Migration completed in PR #456",
    tenant_id="my-app",
)

Use cases for manual linking:

• Correcting missed connections
• Adding domain-specific relationships (e.g., "competes_with", "regulates")
• Creating high-confidence anchors

Operation Comparison

| Operation | Write/Read | LLM Required | Use Case | |---|---|---|---| | remember | Write | No | Store new information | | recall | Read | No | Retrieve relevant context | | forget | Write | No | Remove from results | | search | Read | No | Filtered episode search | | link | Write | No | Manual relationships |

Next Steps