Knowledge Memory

Core Concepts

• Memories — Individual knowledge items with a kind (e.g. pattern, decision, warning), title, body content, and scope path.
• Scoped access — Memories are scoped by path, allowing for project-level or module-level visibility.
• Semantic search — Queries return the most relevant memories based on LLM-powered embedding similarity, not exact keyword matches.
• Proposed → Accepted flow — Memories start as proposed and can be reviewed and accepted, ensuring quality control.

Key Capabilities

• Save eternal truths — Patterns, decisions, and warnings that persist across sessions and agent instances.
• LLM-powered embedding — Semantic search with vector similarity for intelligent retrieval.
• Scope control — Memories are organized by path, giving precise control over what each agent can access.
• Shared across agents — Any agent can query the memory store and benefit from past learnings.

Cognition Specs

Cognition Specs extend the memory system with structured module documentation. Each module can define its purpose, invariants, workflows, and failure modes as structured guidance packets that agents receive when working in that area.

• Purpose — What the module does and why it exists
• Invariants — Rules that must always hold true
• Workflows — Common operation sequences
• Failure modes — Known issues and their resolutions
• Auto-generated guidance packets for agents when they claim related tasks

MCP Tools