Anchor Engine: Deterministic Semantic Memory for LLMs Running on Phones with <1GB RAM
Key Takeaways
- ▸Anchor Engine provides deterministic, persistent semantic memory for LLMs without embedding drift, cloud dependencies, or black-box similarity matching
- ▸The system is lightweight enough to run on Raspberry Pi or budget hardware with <1GB RAM, enabling true local-first AI without API calls
- ▸Uses graph-based retrieval instead of vector search, making results inspectable and traceable with guaranteed consistency across queries
Summary
Anchor Engine is a new open-source memory layer designed to address a fundamental limitation of large language models: their inability to retain information between conversations. Rather than relying on vector embeddings or cloud-based retrieval systems, Anchor Engine uses deterministic graph traversal to create persistent, queryable semantic memory that operates entirely offline on edge devices. The system can run on resource-constrained hardware like Raspberry Pi or budget mini PCs with less than 1GB of RAM, making it practical for local-first AI applications.
The engine works by structuring text into a lightweight graph of concepts and relationships, then uses the STAR algorithm (Semantic Traversal And Associative Retrieval) to retrieve contextually relevant information with complete transparency—users can trace exactly why any piece of information was retrieved. Built on PGlite (WASM-based PostgreSQL), Anchor Engine eliminates native compilation requirements and provides zero-lock-in through an AGPL 3.0 open-source license. Production benchmarks show sub-200ms query latency on datasets containing 25M tokens, with restore speeds of 340 atoms per second.
- Built on PGlite with zero-compilation deployment, cross-platform compatibility (ARM64, x64, Linux, macOS), and AGPL 3.0 open-source licensing to prevent vendor lock-in
Editorial Opinion
Anchor Engine addresses a critical pain point in modern AI applications—the stateless nature of LLMs and the opacity of vector-based retrieval systems. By combining deterministic graph traversal with lightweight, offline-first architecture, it offers a compelling alternative to centralized retrieval systems while maintaining transparency and auditability. The choice to use WASM-based PostgreSQL and open-source licensing suggests genuine commitment to accessibility, though production adoption will depend on how well the graph extraction and traversal algorithms handle diverse real-world data without the flexibility that probabilistic embeddings provide.
