Necessity IS the Mother of Invention
Every Claude session starts from zero. No memory of what you worked on yesterday, no awareness of your project structure, no continuity. If you're doing serious work — writing, engineering, research — you spend the first 10 minutes of every conversation re-explaining who you are and what you're building.
I got tired of it, so I built BOND.
What it does:
Persistent memory across sessions using QAIS (Quantum Approximate Identity Substrate) — a hyperdimensional computing system using 4096-bit bipolar vectors. No neural network, no embeddings API, no external calls. Same text always produces the same resonance pattern. Deterministic and auditable.
A four-class entity system (doctrine, project, perspective, library) that governs what Claude can access and how it behaves in different contexts. Each class has hard tool boundaries — a doctrine entity gets filesystem + semantic analysis, a perspective entity gets memory + growth tools. Cross-class access is structurally forbidden, not just discouraged.
A vine lifecycle for perspective entities inspired by John 15. Seeds auto-collect during conversation when resonance exceeds a threshold. Pruning is the only deliberate act — the perspective evaluates its own growth through its identity lens and decides what stays. No human approval gate on collection. Quality lives downstream.
Session crystallization that routes based on context. Working inside a perspective? Crystal writes to a local field. Working unscoped? Crystal writes to global. Two separate .npz files per perspective — seed field for identity, crystal field for narrative continuity. They never mix.
A visual control panel (React + Node) with entity cards, module status, doctrine viewer, and a clipboard bridge to Claude via AutoHotkey. Every command is a button. The panel is the cockpit, Claude is the engine.
SLA (Spectral Linguistic Addressing) — a deterministic retrieval engine that powers Warm Restore. When you need context from 30 sessions ago, SLA ranks archived handoffs using spectral scoring and returns results with confidence badges.
What it looks like in practice:
Paste one line into PowerShell:
irm https://moneyjarrod.github.io/BOND/install.ps1 | iex
BOND installs, the panel opens at localhost:3000. Add the skill file to a Claude Project, configure two MCP servers (QAIS + ISS), type {Sync}, and Claude picks up where you left off. Full restore from cold boot takes one command.
The technical stack:
React panel + Express sidecar (entity management, WebSocket live updates)
Python MCP servers (QAIS memory, ISS semantic force analysis)
AutoHotkey clipboard bridge (no HTTP polling — event-driven, zero overhead)
NumPy for vector operations (no ML dependencies, no GPU required)
Files as source of truth (markdown doctrine, JSON state, .npz fields)
Design philosophy:
BOND follows a truth hierarchy: code outranks prose, prose outranks memory. State is derived, not stored redundantly. Every write requires both operators to agree. The counter system tracks conversation depth and signals when context is degrading. When prose and code disagree, code wins — unless it's a bug, in which case code gets fixed to match doctrine.
There's a constitutional doctrine (BOND_MASTER) that sits above all entities and defines what the system IS. Any entity's behavior can be audited against those IS statements. New capabilities trigger mandatory doctrine review — the constitution can't be silently outgrown by the system it governs.
What it costs: Nothing. MIT license. Everything is on GitHub.
Why I built it: I'm not a developer by trade. I design systems — calendars, memory architectures, collaboration frameworks. I kept hitting the same wall: Claude is incredibly capable but has no continuity. Every session is a clean slate. BOND exists because I needed it, and I figured other people do too.
It's v1.5.0 — stable, functional, documented. A visual guide with annotated screenshots ships with the repo so you can see what you're getting before installing. Bugs likely exist and will be fixed. Updates are regular — this was built across 116 sessions and counting. New features are still being implemented as the architecture reveals what it needs. The core works today and it's getting tighter.
Happy to answer questions about the architecture, the memory model, or why I went with hyperdimensional computing over embeddings.
Necessity IS the Mother of Invention Every Claude session starts from zero. No memory of what you worked on yesterday, no awareness of your project structure, no continuity. If you're doing serious work — writing, engineering, research — you spend the first 10 minutes of every conversation re-explaining who you are and what you're building. I got tired of it, so I built BOND. What it does:
Persistent memory across sessions using QAIS (Quantum Approximate Identity Substrate) — a hyperdimensional computing system using 4096-bit bipolar vectors. No neural network, no embeddings API, no external calls. Same text always produces the same resonance pattern. Deterministic and auditable. A four-class entity system (doctrine, project, perspective, library) that governs what Claude can access and how it behaves in different contexts. Each class has hard tool boundaries — a doctrine entity gets filesystem + semantic analysis, a perspective entity gets memory + growth tools. Cross-class access is structurally forbidden, not just discouraged. A vine lifecycle for perspective entities inspired by John 15. Seeds auto-collect during conversation when resonance exceeds a threshold. Pruning is the only deliberate act — the perspective evaluates its own growth through its identity lens and decides what stays. No human approval gate on collection. Quality lives downstream. Session crystallization that routes based on context. Working inside a perspective? Crystal writes to a local field. Working unscoped? Crystal writes to global. Two separate .npz files per perspective — seed field for identity, crystal field for narrative continuity. They never mix. A visual control panel (React + Node) with entity cards, module status, doctrine viewer, and a clipboard bridge to Claude via AutoHotkey. Every command is a button. The panel is the cockpit, Claude is the engine. SLA (Spectral Linguistic Addressing) — a deterministic retrieval engine that powers Warm Restore. When you need context from 30 sessions ago, SLA ranks archived handoffs using spectral scoring and returns results with confidence badges.
What it looks like in practice: Paste one line into PowerShell: irm https://moneyjarrod.github.io/BOND/install.ps1 | iex BOND installs, the panel opens at localhost:3000. Add the skill file to a Claude Project, configure two MCP servers (QAIS + ISS), type {Sync}, and Claude picks up where you left off. Full restore from cold boot takes one command. The technical stack:
React panel + Express sidecar (entity management, WebSocket live updates) Python MCP servers (QAIS memory, ISS semantic force analysis) AutoHotkey clipboard bridge (no HTTP polling — event-driven, zero overhead) NumPy for vector operations (no ML dependencies, no GPU required) Files as source of truth (markdown doctrine, JSON state, .npz fields)
Design philosophy: BOND follows a truth hierarchy: code outranks prose, prose outranks memory. State is derived, not stored redundantly. Every write requires both operators to agree. The counter system tracks conversation depth and signals when context is degrading. When prose and code disagree, code wins — unless it's a bug, in which case code gets fixed to match doctrine. There's a constitutional doctrine (BOND_MASTER) that sits above all entities and defines what the system IS. Any entity's behavior can be audited against those IS statements. New capabilities trigger mandatory doctrine review — the constitution can't be silently outgrown by the system it governs. What it costs: Nothing. MIT license. Everything is on GitHub. Why I built it: I'm not a developer by trade. I design systems — calendars, memory architectures, collaboration frameworks. I kept hitting the same wall: Claude is incredibly capable but has no continuity. Every session is a clean slate. BOND exists because I needed it, and I figured other people do too. It's v1.5.0 — stable, functional, documented. A visual guide with annotated screenshots ships with the repo so you can see what you're getting before installing. Bugs likely exist and will be fixed. Updates are regular — this was built across 116 sessions and counting. New features are still being implemented as the architecture reveals what it needs. The core works today and it's getting tighter. Happy to answer questions about the architecture, the memory model, or why I went with hyperdimensional computing over embeddings.