Skip to main content

​
How it works

Episodic memories (from conversations) and static knowledge (from documents) have different lifecycles. MemWire keeps them separate but surfaces both at recall time. When you call memory.recall(), the engine:
  1. Traverses the memory graph for relevant conversation paths
  2. Also searches knowledge chunks using the same query embedding
  3. Merges both results into a single RecallResult
You can also search knowledge directly without going through the memory graph.

​
Adding a knowledge base

from memwire import MemWire, MemWireConfig

config = MemWireConfig(qdrant_path="./memwire_data")
memory = MemWire(config=config)

USER_ID = "alice"

memory.add_knowledge(
    name="Company Handbook",
    chunks=[
        {"content": "All employees must complete onboarding within the first 30 days."},
        {"content": "Remote work is allowed up to 3 days per week."},
        {"content": "Expenses above $500 require manager approval."},
    ],
    user_id=USER_ID,
)
Chunks are embedded and stored in a dedicated vector store collection, separate from memory vectors.

​
Recall with knowledge

Knowledge chunks appear automatically in recall results: 
result = memory.recall("What is the expense approval policy?", user_id=USER_ID)

# Conversation memories
print(result.formatted)

# Knowledge chunks (also included in formatted)
for chunk in result.knowledge:
    print(chunk.content, chunk.score)

chunks = memory.search_knowledge(
    "remote work policy",
    user_id=USER_ID,
    top_k=5,
)
for chunk in chunks:
    print(f"[{chunk.score:.2f}] {chunk.content}")

​
Deleting a knowledge base

# List knowledge bases via stats
stats = memory.get_stats(user_id=USER_ID)
print(stats["knowledge_bases"])  # count

# Delete by kb_id
memory.delete_knowledge(kb_id="kb_abc123")

​
Scoping knowledge

Knowledge bases respect the same multi-tenancy hierarchy as memories: 
memory.add_knowledge(
    name="Team Runbook",
    chunks=[{"content": "Deploy with: ./scripts/deploy.sh production"}],
    user_id=USER_ID,
    workspace_id="team_engineering",
    app_id="ops_assistant",
)

​
KnowledgeChunk fields

FieldTypeDescription
chunk_idstrUnique identifier for the chunk
kb_idstrParent knowledge base ID
contentstrChunk text
scorefloatSimilarity score from the search
metadatadictOptional metadata passed at ingestion