@loreai/core

Experimental — Under active development. APIs, storage format, and behavior may change.

Shared memory engine for Lore — three-tier storage, distillation, curation, gradient context management, and FTS5-backed recall.

This package is host-agnostic. It doesn't ship a user-facing extension on its own; it's consumed by adapter packages that wire it into a specific coding agent:

• @loreai/opencode — OpenCode plugin (also published as opencode-lore legacy alias)
• @loreai/pi — Pi coding-agent extension

Install

npm install @loreai/core
# or
pnpm add @loreai/core

You only need to install this directly if you're building a new adapter. End users install one of the host packages above.

Vector embeddings

Recall uses @huggingface/transformers with nomic-embed-text-v1.5 (768-dim INT8 quantized, ~137 MB) for on-device vector search — no API key required.

• Binary mode (standalone lore binary): Uses WASM backend (onnxruntime-web). Works on all platforms without native dependencies.
• npm mode (npm install @loreai/gateway): Uses native backend (onnxruntime-node). May fail on some configurations (e.g. CUDA 13 on Linux/x64 — microsoft/onnxruntime#26586).

When local embeddings aren't available, recall has graceful fallbacks:

1. Remote auto-fallback — set VOYAGE_API_KEY or OPENAI_API_KEY in your env. The first embed() call detects the missing local provider and swaps over for the rest of the process. Voyage wins ties.
2. FTS-only — if no remote keys are set and local embeddings fail, recall uses SQLite FTS5 full-text search. Still functional, just without vector similarity ranking.

To pin a specific provider, set search.embeddings.provider in .lore.json:

{ "search": { "embeddings": { "provider": "voyage" } } }

(also supports "openai"; reads VOYAGE_API_KEY / OPENAI_API_KEY from env).

Documentation

Full architecture, benchmarks, and rationale: github.com/BYK/loreai

License

FSL-1.1-Apache-2.0 — see LICENSE.