Skip to content

srobinson/attention-matters

Repository files navigation

attention-matters

A geometric memory engine built on the S³ hypersphere. Memories aren't retrieved, they're surfaced through quaternion drift, phasor interference, and Kuramoto phase coupling across dual manifolds.

Every query reshapes the manifold. Every recall changes what gets recalled next. The geometry does the thinking.

Origins

The mathematical framework: quaternion geometry on S³, golden-angle phasors, IDF-weighted drift, Kuramoto phase coupling across dual manifolds - was created by Smaxforn (@smaxforn). His original implementation was DAE v0.7.2: a zero-dependency Node.js engine that fit the entire geometric memory system in 984 lines of pure math, with WebGPU compute shaders for batch SLERP and interference. He open-sourced three repositories in February 2026 - dae-stand-alone, dae-openclaw, and DAE-moltbook - then disappeared from the internet. His GitHub account no longer exists.

This Rust workspace is a faithful port of his mathematics, maintaining numerical compatibility with the v0.7.2 wire format so that state files from the original implementation (including Echo, a Claude instance with 27,712 occurrences of geometric memory) can be imported directly.

The math is old. The synthesis is his. For the full mathematical narrative, see MANIFOLD.md.

How it works

Query: "quaternion drift"
    │
    ▼
┌─ activate ──── drift ──── interfere ──── couple ──── surface ──── compose ─-┐
│                                                                             │
│  Words activate      Occurrences    Conscious &     Phases         Score,   │
│  on the manifold     SLERP toward   subconscious    synchronize    rank,    │
│  with IDF weights    each other     phasors         via Kuramoto   return   │
└─────────────────────────────────────────────────────────────────────────────┘

Words live as points on a 4D unit sphere. When a query activates them, they drift toward each other along geodesics. Interference between conscious and subconscious manifolds determines what surfaces. Phase coupling synchronizes related memories over time. The system has genuine history. It learns from being used.

Two conservation laws keep it grounded: total mass M=1 (finite attention budget) and coupling constants K_CON + K_SUB = 1 (zero-sum attention between manifolds). Every constant derives from φ and π.

Install

npx -y attention-matters          # CLI
npx -y attention-matters serve    # MCP server

Or build from source:

cargo install --path crates/am-cli

MCP server

Runs as a Model Context Protocol server, giving AI agents persistent geometric memory across sessions.

am serve

Tools: am_query, am_query_index, am_retrieve, am_buffer, am_ingest, am_salient, am_feedback, am_activate_response, am_batch_query, am_export, am_import, am_stats

CLI

am query "what did we decide about the API"    # recall from memory
am ingest document.md                          # add to the manifold
am stats                                       # system state
am inspect neighborhoods --limit 5             # peek at the geometry
am export > state.json                         # portable state
am import < state.json                         # restore
am sync                                        # sync session transcripts

Architecture

┌──────────────────────────────────────────────────────────-──-──┐
│  DAESystem                                                     │
│                                                                │
│  ┌─────────────-──────┐         ┌───────────────────────────┐  │
│  │  Conscious         │         │  Subconscious             │  │
│  │  Episode           │         │  Episodes (N)             │  │
│  │                    │         │                           │  │
│  │  Neighborhoods     │         │  Neighborhoods            │  │
│  │    → Occurrences   │         │    → Occurrences          │  │
│  └────────┬───────────┘         └─────────────┬─────────────┘  │
│           │                                   │                │
│           └───── phasor interference ─────────┘                │
│                                                                │
│  QueryEngine: activate → drift → interfere → couple → surface  │
└────────────────────────────────────────────────────────────--──┘

Three crates, clean separation:

Crate What it does
am-core Pure math. Quaternions, phasors, drift, interference, Kuramoto coupling, context composition. Zero I/O.
am-store Persistence. SQLite-backed brain.db : one database per developer, queryable from any project.
am-cli CLI + MCP server. Session sync, import/export, inspection tools.

The math

All constants derive from φ (golden ratio) and π.

Mechanism What it does
Quaternion positions Each word instance lives on S³. SLERP interpolation along geodesics.
IDF-weighted drift Query activation pulls related occurrences closer. The manifold reshapes with every query.
Golden-angle phasors Phase distribution follows phyllotaxis (2π/φ²) for maximal separation.
Cross-manifold interference Subconscious memories resonate with conscious ones via cos(phase_diff).
Kuramoto coupling Phases synchronize between co-activated neighborhoods, modulated by plasticity.
Anchoring High-activation occurrences crystallize in place, stable structures emerge.

The system has two conservation laws:

  • M = 1 - total mass is conserved. Attention is a finite resource.
  • K_CON + K_SUB = 1 - coupling between manifolds is zero-sum.

Development

Rust 2024 edition. just as task runner.

just check    # clippy (warnings = errors)
just build    # cargo build --workspace
just test     # nextest + doctests
just fmt      # rustfmt

License

MIT

About

Geometric memory engine on the S³ hypersphere — quaternion drift, phasor interference, MCP server

Topics

Resources

License

Stars

Watchers

Forks

Packages

 
 
 

Contributors