Anchor Engine ⚓

Deterministic semantic memory for local‑first AI systems

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.

---

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.

---

What Anchor Engine Is

A semantic memory layer that provides:

• Deterministic retrieval
• Graph‑based semantics
• Temporal decay
• Provenance receipts
• CPU‑only performance (<1GB RAM)
• Local‑first architecture

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

• ❌ An agent framework
• ❌ A cloud service
• ❌ A probabilistic vector search engine
• ❌ A tool that stores your data anywhere except your machine

---

Quick Start

git clone https://github.com/RSBalchII/anchor-engine-node
cd anchor-engine-node
./install.sh   # or install-macos.sh / install.ps1
pnpm start

Open: http://localhost

---

How It Works (High‑Level)

1. Content lives on disk, not in the database

The database stores pointers only:

• file path
• byte offsets
• tags
• timestamps
• provenance

All content is read directly from disk.
This keeps RAM low and makes the DB disposable.

2. Retrieval uses the STAR algorithm

STAR = Semantic Temporal Associative Retrieval

It combines:

• Semantic gravity (shared tags × hop distance)
• Temporal decay (recent memories pull harder)
• Structural gravity (SimHash proximity)

The scoring equation is deterministic and explainable.
Every result includes a receipt showing why it was retrieved.

3. The engine rebuilds itself on startup

Your data lives in:

• notebook/inbox/ (trusted, high‑weight)
• notebook/external-inbox/ (external, normal weight)

The engine mirrors and indexes these on launch.

---

MCP Integration

Anchor Engine can run as an MCP server, giving any MCP‑compatible agent persistent memory.

Available tools:

• /anchor_query — semantic search
• /anchor_distill — session distillation
• /anchor_illuminate — graph traversal
• /anchor_read_file — byte‑range file reads
• /anchor_ingest_text — opt‑in write operations

Works with:

• Claude Desktop
• Cursor
• Qwen Code

---

Example Query

curl -X POST http://localhost/v1/memory/search \
  -H "Content-Type: application/json" \
  -d '{
    "query": "What did we decide about OAuth?",
    "token_budget": 2048
  }'

Returns:

• matched atoms
• provenance
• hop distance
• temporal weighting
• byte‑range pointers

---

Security Posture

Anchor Engine is designed for local‑first, offline use.

Security features include:

• Path traversal prevention
• API key authentication
• Rate limiting
• Input validation
• Zero-copy deduplication
• Disposable ephemeral DB
• No cloud dependencies

A full CodeQL audit (April 2026) found no unmitigated high‑severity issues.

---

Benchmarks

Dataset: 91MB chat history (~25M tokens)

• Ingestion: 178 seconds
• Search p95: <200ms
• Idle memory: ~600MB
• Restore: 281k atoms in 13.8 minutes

Hardware: consumer laptop, no GPU.

---

Why “Anchor”?

Drop a query into your memory graph.
It sinks to the semantic bottom.
The chain plays out to your chosen radius.
What you retrieve is what’s relevant — nothing more, nothing less.

Same anchor, same spot, same result.

---

Documentation

• Quick Start
• API Reference
• MCP Integration
• Architecture Overview
• STAR Algorithm Whitepaper
• Security Standards
• Benchmarks

---

License

AGPL‑3.0

---

Citation

@software{anchor_engine,
  title={STAR: Semantic Temporal Associative Retrieval},
  author={Balch II, R.S.},
  version={5.0.0},
  year={2026},
  url={https://github.com/RSBalchII/anchor-engine-node}
}