Anchor Engine ⚓

June 23, 2026 · View on GitHub

Deterministic semantic memory for local‑first AI systems

⚠️ FIRST READ THIS: The Documentation Policy is the single source of truth for all abstractions, file locations, and developer conventions. All runtime objects are stored in $HOME/.anchor/ and are never in the project root. Read it before coding.

👨‍💻 Developers - Building, Testing, Debugging

Start here:

Documentation Policy - Required read (abstractions, paths, conventions)
Project Specs - Architecture, API, data model, test framework
Current Standards - Active architecture standards (001-038)

Quick reference:

Test Framework - How to test (P0, Vitest, integration, operational)
Test Usability - Automated runtime verification

🧠 Users - Understanding, Using, Exploring

Start here:

Whitepaper - The STAR Context Protocol (conceptual, no code)
API Endpoints - Available API routes and usage
Architecture Overview - How it works (high-level)

For deeper knowledge:

STAR Algorithm - The physics-inspired retrieval engine
Data Model - Compound → Molecule → Atom hierarchy

📚 Documentation Structure

Directory	Audience	Purpose	Key Files
`specs/`	🧑‍💻 Developers	Technical architecture, implementation, API	spec.md, current-standards/
`docs/`	🧠 Users	Conceptual understanding, whitepaper, theory	whitepaper.md
`engine/`	🧑‍💻 Developers	Source code, tests, API routes	src/, tests/

🚀 Quick Start

Installation (All Users)

# Clone and install
git clone https://github.com/RSBalchII/anchor-engine-node
cd anchor-engine-node
pnpm install

# Start the engine
pnpm start

Development (Developers)

# Run tests
pnpm test

# Start with logging
pnpm start

# Run operational verification
python tests/test_us006.py

Windows 11 Startup Notes ⚠️

On Windows, pnpm.cmd is a PowerShell wrapper that can cause startup issues in git-bash/MSYS environments. If you encounter errors like:

ERR_MODULE_NOT_FOUND: Cannot find module 'C:\nvm4w\nodejs\node_modules\pnpm\dist\pnpm.mjs'
pnpm: command not found
Engine fails to start via pnpm start

Use this workaround instead:

# Direct Node.js startup (bypasses pnpm wrapper)
node --expose-gc engine/dist/index.js

This approach:

✅ Works in git-bash/MSYS without path translation issues
✅ No dependency on nvm4w or global pnpm installation quirks
✅ More reliable than pnpm start on Windows 10/11

Alternative methods:

# If you have pnpm installed globally
C:\Users\rober\AppData\Roaming\npm\npx.cmd pnpm start

# Or use the PowerShell wrapper directly
powershell -NoProfile -ExecutionPolicy Bypass -File "pnpm.ps1" start

Note: The compiled engine/dist/index.js is built via pnpm --filter anchor-engine build. If you need to rebuild after code changes, run:

C:\Users\rober\AppData\Roaming\npm\node_modules\pnpm\bin\pnpm.mjs --filter anchor-engine build

📖 What Anchor Engine Is

Anchor Engine is a semantic memory layer — not an agent framework, not a vector database, and not a cloud service.

It's a deterministic, explainable, CPU‑only system for storing and retrieving long-term memory for AI agents.

It replaces embeddings with a physics‑inspired graph algorithm (STAR) that retrieves context using structure, time, and meaning — not dense vectors.

What Anchor Engine Provides

✅ Deterministic retrieval — same query → same result
✅ Graph‑based semantics — not probabilistic similarity
✅ Temporal decay — older memories naturally fade
✅ Provenance receipts — every retrieval comes with proof
✅ CPU‑only performance — <1GB RAM, no GPU needed
✅ Local‑first architecture — your data never leaves your machine

Use Cases

Use it as:

A drop‑in replacement for embeddings/vector DBs
A memory backend for MCP‑compatible agents (Claude, Cursor, Qwen Code)
A personal knowledge system for long‑term projects

What Anchor Engine Is Not

❌ Not an agent framework
❌ Not a cloud service
❌ Not a probabilistic vector search engine
❌ Not a tool that stores your data anywhere except your machine

Why This Exists

Most AI memory systems today assume:

GPU‑heavy vector search
Probabilistic retrieval
Cloud dependence
Opaque similarity metrics

I kept running into the same problem:

I needed a memory system that behaved like a mind, not a search engine.

Anchor Engine is built around three principles:

Determinism — same query → same result
Explainability — every retrieval comes with provenance
Local sovereignty — your data stays on your machine

This project grew out of months of building agents that needed reliable memory — and discovering that existing tools weren't designed for that job.

🎓 Citation

If you use this software in your research, please cite:

DOI: https://doi.org/10.5281/zenodo.19324840
Citation: Balch II, R. S. (2026). STAR: Semantic Temporal Associative Retrieval - A Local-First Graph-Based Context Engine (v5.0.0). Zenodo.

Software: Anchor Engine Node
License: AGPL-3.0

⚡ Architecture & Technology Stack

Native Modules (Rust WASM)

Anchor Engine uses Rust-compiled WebAssembly modules for performance-critical operations. This eliminates the need for native compilation and provides universal platform support.

Published Packages:

Package	Purpose	Version
@rbalchii/anchor-fingerprint-wasm	Content fingerprinting (MD5, SHA256)	1.0.0+
@rbalchii/anchor-atomizer-wasm	Text atomization & entity extraction	1.0.0+
@rbalchii/anchor-keyextract-wasm	Key-value extraction from text	1.0.0+
@rbalchii/anchor-tagwalker-wasm	Semantic tag traversal	1.0.0+

Benefits:

✅ Zero native compilation required (works on Windows ARM64, macOS, Linux)
✅ ~90% smaller binary size (~1.4 MB WASM vs ~12 MB C++ DLLs)
✅ 8x faster module loading (benchmarks: ~50ms vs ~400ms for native)
✅ Universal platform support

Note: The older C++ native modules (engine/src/native/ directory) have been deprecated and removed in favor of these Rust WASM packages. See engine/src/README.md for full architectural details.

QwenPaw Agent Configuration Protection 🔒

To protect sensitive QwenPaw agent configuration files from being accidentally committed to the repository, the following files are now gitignored:

BOOTSTRAP.md
MEMORY.md
PROFILE.md
SOUL.md
AGENTS.md

These files may contain private information and should never be shared. The .gitignore has been updated to prevent accidental commits of these sensitive configuration files.