Pi with Lore

Pi is a terminal-based AI coding agent. The @loreai/pi extension loads inside Pi and routes every LLM call through the Lore gateway.

Install

Add the extension to your ~/.pi/settings.json:

{
  "packages": [
    "npm:@loreai/pi@latest"
  ]
}

Then run pi install once. The extension auto-loads on every Pi session.

What you get

Once the extension is loaded, every conversation is captured in Lore’s three-tier memory. Distillations run in the background, the recall tool is available, and your project knowledge is exported to .lore.md and AGENTS.md automatically. See the architecture overview for the full picture.

Local embeddings

Recall uses @huggingface/transformers with nomic-embed-text-v1.5 (768-dim INT8 quantized, ~137 MB) for on-device vector search by default — no API key required. The model downloads on first use and is cached locally.

If local embeddings fail (for example, CUDA 13 on Linux/x64 with onnxruntime-node), set VOYAGE_API_KEY or OPENAI_API_KEY in your environment and recall transparently switches to that provider. You can also pin a hosted provider explicitly in .lore.json:

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

If none of the above apply, recall falls back to FTS-only search.

Pointing Pi at a local LLM server

The Pi plugin reads LORE_UPSTREAM_<PROVIDER> from your environment and injects it as the x-lore-upstream-url header on each request. Set the env var for the provider you want:

export LORE_UPSTREAM_VLLM=http://localhost:8000
# or
export LORE_UPSTREAM_OLLAMA=http://localhost:11434
# or
export LORE_UPSTREAM_LLAMACPP=http://localhost:8080
# or
export LORE_UPSTREAM_LMSTUDIO=http://localhost:1234
# or
export LORE_UPSTREAM_TGI=http://localhost:8080
# or
export LORE_UPSTREAM_LITELLM=http://localhost:4000

The URL is the server root — do not include /v1 (the gateway appends API paths automatically). See the local inference guide for a full walkthrough.

Per-harness notes

• Compaction is intercepted. Pi’s compaction goes through its extension API rather than HTTP. The Pi extension calls the gateway’s POST /v1/compact endpoint to get a full LLM-synthesized compaction summary (force-distill + knowledge + compact prompt).
• Provider headers are auto-injected. The extension’s registerProviders() walks the configured providers and sets the gateway base URL, x-lore-session-id, x-lore-project, and x-lore-git-remote headers on each request.
• LORE_GATEWAY_URL overrides auto-detection. If the gateway is on a non-default host or port, set LORE_GATEWAY_URL in your environment before launching Pi.

Next steps

• Architecture — how temporal storage, distillation, and the gradient context manager fit together.
• Configuration — full reference for .lore.json.
• Local inference — running Pi against Ollama, vLLM, or llama.cpp.
• Custom upstreams — corporate proxies, LiteLLM, Cloudflare AI Gateway.