This repository contains a technical proposal for a missing layer in current LLM / chat-based architectures:
a marker–vector memory graph that enables long-horizon, multi-branch reasoning with stable user and task state.
The documents are written for three primary audiences:
- Product / Leadership – why this is a P0 architectural gap with direct impact on UX, trust and revenue.
- System Architects / Platform Engineers – how to integrate a marker–vector memory layer into existing stacks.
- ML Researchers – how to formulate this as a structured memory / latent state problem aligned with AGI-oriented research.
The goal is not to replace existing transformer-based models, but to add a missing substrate that:
- Provides stable long-lived preferences (account-level).
- Tracks projects and tasks across sessions (project-level).
- Structures per-dialog reasoning as a graph of semantic markers (session-level).
- Aligns monetization with resource prioritization (QoS), instead of feature gating.
docs/01_executive_summary.md– high-level overview and business framing (P0 gap).docs/02_architecture_proposal.md– systems-level architecture for the marker–vector memory graph.docs/03_research_spec.md– research-facing specification and directions for experiments.meta/change_log.md– version history of this proposal.
This repository is licensed under the MIT License (LICENSE file).
It is intentionally permissive to allow free use, modification and integration of these ideas into existing systems and research.
Version: v0.1 (initial draft)
Scope: Problem framing, architecture outline, research directions, and monetization alignment.