BrainBox v5.0 — Hebbian Muscle Memory for AI Agents

What's New in v5.0

BrainBox v5.0 is the most complete Hebbian memory system for AI coding agents. Zero vectors. Zero RAG. Pure co-access learning.

New Features

• Session Intent Capture — records what you were trying to do, not just which files you touched
• Hub Detection — identifies the most-connected neurons in your network (brainbox hubs)
• Staleness Alerts — detects decaying superhighways before they fade (brainbox stale)
• Project Tagging — auto-scopes recall to your current project, reducing cross-project noise
• Anti-Recall Escalation — files recalled but never opened get progressively stronger decay (compound: 10% → 19% → 27%)
• Session Replay — full agent observability for debugging memory behavior
• Activity Log — continuous file access logging for VaultGraph integration

Core

• 59 tests, all passing
• SNAP sigmoid plasticity with BCM sliding threshold
• 2-hop spreading activation with fan-effect normalization
• Error-fix immune system — remembers which files fixed which errors
• Tool sequence myelination — learns your Grep→Read→Edit chains
• macOS daemon (opt-in) — learns from VS Code, Xcode, vim, shell via FSEvents

Integrations

• Claude Code — auto-installs hooks + MCP server on npm install
• Kilo/OpenCode — native plugin
• OpenClaw — NeuroVault memory slot plugin
• Any MCP agent — standalone MCP server with 6 tools

Install

npm install brainbox-hebbian

Paper

Full algorithm details in WHITEPAPER.md — covers SNAP plasticity, BCM theory, spreading activation, error-fix learning, and production benchmarks (67% top-1 recall, 8.9% token savings).

---

Not a vector database. Not RAG. Muscle memory.