Skip to content

overview.md

Latest commit

 

History

History
74 lines (51 loc) · 2.47 KB

overview.md

File metadata and controls

74 lines (51 loc) · 2.47 KB

Cortex Overview

Remram Cortex is the knowledge authority layer for Remram.

It coordinates policy, hot working continuity, durable semantic memory, operational knowledge, and source-of-record evidence without collapsing them into one storage model.

The architecture is now centered on five layers:

  1. Policy
  2. Working Memory
  3. Durable Memory
  4. Operational Knowledge
  5. Evidence

The clean authority model is:

  • Layer 1 = behavioral truth
  • Layer 2 = hot working continuity
  • Layer 3 = durable semantic truth
  • Layer 4 = operational knowledge truth
  • Layer 5 = source-of-record evidence

Chosen Stack

The active stack direction is:

  • OpenClaw as the chosen agentic framework
  • Cortex-owned Layer 1 policy
  • QMD for Layer 2 hot working memory and notions
  • a narrow High-Signal Mamba stream as a later always-on supercharge layer
  • Graphiti + Neo4j for Layer 3 durable memory
  • Postgres for the operational middle of the stack
  • a Layer 5 evidence system, with runtime_evidence, reference_cache, and authored_artifact as the main source-of-record classes

What Cortex Owns

Cortex owns:

  • policy composition
  • Layer 2 notion and high-signal lifecycle rules
  • Layer 3 durable-memory contracts
  • Layer 4 operational knowledge authority rules
  • the split between operational knowledge and source-of-record evidence
  • reflection, Dream, reconciliation, and promotion behavior

OpenClaw Principle

OpenClaw was not chosen because Cortex benchmarked orchestrators and decided it was the abstract best shell.

It was chosen because it is the framework Remram intends to build around.

That means the architectural rule is:

  • do things the OpenClaw way when that path is good enough
  • extend OpenClaw cleanly before replacing it
  • add memory and knowledge behavior around the framework boundary

What Changed In The Locked Design

The biggest clarifications now are:

  • Layer 2 explicitly uses QMD
  • Mamba is narrow by design
  • Mamba is not required to make the architecture valid and lands later as a Phase 5 supercharge
  • Layer 3 stays one Graphiti memory system
  • Layer 4 is the operational knowledge authority
  • Layer 5 is the evidence layer, with authored canon as a permanent subset rather than the definition of the layer
  • OpenSearch is not part of the near-term stack

Repository Direction

The repository now aims to describe one coherent architecture rather than a stack-comparison process.

Historical alternatives and superseded packages live in the top-level archive.