@openweave/weave-link

🧠 weave-graph

WeaveGraph — The knowledge graph engine at the heart of OpenWeave.

Part of the OpenWeave monorepo.

What it does

WeaveGraph manages all memory for an OpenWeave session:

• Stores concepts, decisions, errors and corrections as typed graph nodes
• Connects them with semantic edges (relates, causes, corrects, implements, depends_on)
• Compresses context into the graph when the LLM window reaches 75% capacity
• Retrieves relevant nodes from long-term memory given a query
• Persists everything to disk by chat_id, survives across sessions

Node Types

Type	Description
CONCEPT	A key idea or term in the project
DECISION	An architectural or implementation decision
MILESTONE	A planned deliverable
ERROR	A response flagged as incorrect by the user (suppressed)
CORRECTION	The correct version linked to an ERROR node
CODE_ENTITY	A function, class, or module created during the session

Edge Types

Relation	Meaning
RELATES	General semantic relationship
CAUSES	A causes B
CORRECTS	A is the correction of B (B is an ERROR node)
IMPLEMENTS	A implements B (code → decision)
DEPENDS_ON	A depends on B
BLOCKS	A blocks B

Quick Start

from weave_graph import ContextGraphManager, NodeType, EdgeType

# Initialize for a session
graph = ContextGraphManager(chat_id="proj_abc123")

# Add a decision node
decision = graph.add_node(
    content="Use PostgreSQL for persistence, not SQLite — project will scale",
    node_type=NodeType.DECISION,
    tags=["database", "architecture"]
)

# Add a concept and relate it
concept = graph.add_node(
    content="Session persistence by chat_id",
    node_type=NodeType.CONCEPT
)
graph.add_edge(decision.id, concept.id, EdgeType.RELATES)

# Suppress an error and record correction
graph.suppress_error_node(
    node_id=bad_response_node.id,
    correction_content="Use async/await, not threading — the codebase is fully async"
)

# Query relevant context before responding
relevant = graph.query_relevant_nodes("database connection pooling", top_k=5)

# Compress context when window is getting full
compressed_summary = graph.compress_context_to_graph(
    context_text=long_conversation_text,
    llm_extractor_fn=my_extraction_function
)

Storage

All data is persisted to:

{storage_path}/{chat_id}/
├── context_graph.json    ← Full graph (nodes + edges)
├── roadmap.md            ← Human-readable milestone status
├── decisions.md          ← Decision log
└── errors.md             ← Error pattern registry

Installation

pip install openweave-graph
# or within the monorepo:
pnpm --filter weave-graph dev