Cortex
Memory that doesn’t lie.
Cortex is a per-actor typed memory graph. Every fact, action, and summary is written once, anchored with a Merkle proof, and never silently overwritten.
Typed memory graph
Nodes in the graph carry declared types: facts, summaries, preferences, credentials, and more. Type-safe retrieval at query time.
Append-only journals
Every write is appended, never mutated. The full revision history is available for any node. Silent data loss is impossible.
Merkle-anchored snapshots
Each snapshot is anchored in a Merkle tree. Any state transition can be cryptographically verified — on-chain or off.
Paged context injection
The Executor retrieves only the relevant pages of memory for each session, preventing context overflow without losing continuity.
Per-actor isolation
Every agent has its own isolated memory namespace. Agents cannot read each other’s memory unless explicitly delegated.
Automatic summarization
Long-running sessions are automatically summarized and compressed, preserving key facts while minimizing token cost.
Built on Pebble
Cortex is powered by Pebble, an append-only key-value store designed for high-throughput agent workloads. Pebble provides the durability and concurrency guarantees that memory-critical agent infrastructure demands.
In production environments, Cortex clusters are sharded by actor ID and replicated for availability. In local and personal deployments, the full stack runs in a single process with no external dependencies.