Skip to content

technology-stack.md

Latest commit

 

History

History
203 lines (131 loc) · 6.23 KB

technology-stack.md

File metadata and controls

203 lines (131 loc) · 6.23 KB

Technology Stack

This document records the active Cortex stack after design lock.

It is not a comparison memo.

It describes the stack the repository is being organized around and the sequencing posture for implementing it.

Long-Term Stack

The long-term stack is:

  • OpenClaw as the chosen agentic framework
  • Cortex-owned Layer 1 policy composition
  • QMD for Layer 2 hot working memory and notions
  • a narrow High-Signal Mamba stream as a later Layer 2 supercharge
  • Graphiti + Neo4j for Layer 3 durable semantic memory
  • Postgres as the operational middle-layer authority
  • a Layer 5 evidence system, with Git as one backend for authored_artifact when publication-grade canon is actually warranted

Delivery Sequencing

The base architecture is complete without live Mamba.

The sequencing choice is:

  • Phase 0 establishes a vanilla OpenClaw baseline with QMD
  • Phase 1 activates the chat-derived Cortex loop
  • Phase 2 adds one-way reference decomposition
  • Phase 3 adds owned-source maintenance and dirty-state semantics
  • Phase 4 adds authored artifact promotion into Git
  • Phase 5 adds Mamba, Intuition, and optimization work

This is a sequencing choice, not an architecture reversal and not a sign that the stack is waiting to become valid later.

Why This Stack

OpenClaw

OpenClaw is the chosen agentic framework for Remram.

That means the selection logic is:

  • build around OpenClaw as the framework center
  • prefer native OpenClaw patterns when they are good enough
  • add Cortex memory and knowledge behavior around that boundary

OpenClaw remains responsible for:

  • runtime execution
  • sessions
  • transcript continuity
  • hooks
  • compaction

Policy In Cortex

Layer 1 remains custom because Cortex-specific behavior lives there.

It owns:

  • role and mode composition
  • approval posture
  • escalation posture
  • prompt-budget discipline
  • mutable preference-policy

Hard runtime and tool-use enforcement should remain as close as possible to OpenClaw agent or plugin configuration.

QMD In Layer 2

QMD is now the explicit Layer 2 hot working-memory substrate.

It is used for:

  • hot working-memory retrieval
  • notion storage
  • fast cross-thread continuity under tighter retrieval rules
  • short-horizon continuity around the OpenClaw session surface

This is not a giant new memory architecture.

It is the pragmatic hot-memory choice inside an OpenClaw-centered stack.

Semantic Processing Before Mamba

Phase 1 still needs semantic processing.

It just does not need an always-on listener yet.

Phase 1 uses:

  • turn-end extraction
  • session-end extraction
  • explicit-checkpoint extraction when needed

Those hooks should survive into later phases.

When Mamba arrives in Phase 5, it augments those hooks instead of replacing them.

High-Signal Mamba Stream Later

The Mamba stream remains narrow by design.

It is:

  • a small always-on listener
  • a typed high-signal channel
  • a Layer 2-adjacent signal producer

It is not:

  • a universal reflection engine
  • the document decomposition engine
  • a general-purpose semantic authority

Its role is to supercharge:

  • infinite-context-like continuity without claiming literal infinite recall
  • near-time continuity compression
  • always-on high-signal capture
  • lower-latency semantic awareness
  • cleaner live-session handling for long-running work

It also gives Reflection and Dream better directional signal and better candidate surfaces.

It does not change the authority model.

Reflection and Dream still retain access to fuller evidence and evidence packages for deeper passes.

When Mamba exists, Cortex can add Intuition on top of that stream.

Intuition is the later signal evaluator that decides when high-signal windows should wake hotter layered ingestion or opportunistic GPU-backed processing.

The important line is:

  • the spine already works without Mamba
  • Mamba improves the quality and timeliness of signal flowing through that same spine
  • deeper reflection, Dream, and reconciliation still use fuller evidence when needed

Graphiti + Neo4j

Layer 3 stays centered on Graphiti + Neo4j because it gives:

  • durable semantic memory
  • temporal lineage
  • support relationships
  • invalidation and supersession
  • concept and identity relationships across threads and workspaces

There is only one Graphiti memory system.

It is used for normal durable-memory behavior, including concept clustering and workspace relationship mapping.

Postgres

Postgres remains the practical operational middle-layer authority.

It covers:

  • policy and control data
  • Layer 4 operational knowledge bodies
  • incubation workspaces
  • reference summaries and links
  • decomposed artifact knowledge where canonical artifacts exist
  • retrieval metadata
  • workflow and dirty-state fields
  • evidence-control and indexing state where needed by the current implementation

This is not being justified as the final winner for document retrieval in the abstract.

It is being chosen because it keeps the near-term service count lower and covers the mixed operational middle of the stack well enough.

Git

Git remains one Layer 5 backend for the authored_artifact evidence class when publication-grade artifacts are warranted.

It should not be treated as the default destination for every useful idea, every transcript, or every external reference.

What Is Not In The Active Stack

The active stack explicitly excludes:

  • OpenSearch
  • a second graph system
  • a second Graphiti usage pattern
  • a separate giant external working-memory service

OpenSearch may remain a future option if artifact retrieval becomes dominant enough to justify it, but it is not part of the current architecture or MVP.

Stack Summary

The active rule is:

  • let OpenClaw own runtime mechanics
  • let Phase 0 prove the vanilla OpenClaw baseline first
  • let QMD own hot working-memory retrieval and notions
  • let Phase 1 prove the chat-derived Cortex spine with boundary-triggered semantic processing
  • let later phases add references, owned sources, and authored canon in order
  • let Mamba arrive last as a narrow supercharge to an already-working design
  • let Graphiti own durable semantic memory
  • let Postgres own the operational middle
  • let Layer 5 own source-of-record evidence, with Git used only for authored canon when that is warranted