Works offline. Zero API cost. Honest recall.
Your notes already hold the answers — they're just hard to reach. Zero grows a local memory over them that retrieves honestly: it tells you what it knows, and exactly how much it doesn't.
Get lichen-core — freeOpen source, MIT licensed. Pay what you want, only after it earns it.
The products
lichen-core — the fact-mesh brain
A Node.js brain for your notes that runs fully offline on any laptop. Point it at a folder — node grow.js <folder> — and it grows a memory that reinforces with use and decays with neglect, like a real one.
- Honest by design: answers come with structured coverage — score, level, gaps — so it says "I don't know" as data instead of hallucinating.
- Retrieval is pure math: hybrid dense + BM25 search fused with RRF. No LLM required to remember.
- Local HTTP API on 127.0.0.1: /ask, /recall, /learn, /outcome, /health, /journal.
- Optional local Ollama improves embeddings and phrasing — never required, never phoned home to.
- Configurable with LICHEN_DATA and OLLAMA_URL. Single-user by design.
openclaw-factmesh — memory for your agents
Makes lichen-core the memory backend for OpenClaw agents. The "goldfish problem" — agents that forget everything between sessions — ends here, without cloud APIs.
- Recalls honest facts into every prompt, token-budgeted so it never floods the context.
- Captures user-stated facts after each turn. Assistant text is never captured, and no LLM extraction decides what to keep.
- Agent tools: memory_search, memory_add, memory_forget.
- CLI:
openclaw factmesh statusandopenclaw factmesh search.
The Local AI Field Manual
Everything we learned running real AI on ordinary hardware — no cloud, no API bill, no PhD. Eight chapters of field-tested setup: picking a model your machine can actually run, Ollama tuning, honest retrieval over your own notes, and growing a local memory that admits what it doesn't know.
- Hardware-to-model matching: what runs on 8GB, 16GB, and a used 24GB card — with real numbers, not marketing.
- Ollama from install to tuned: quantization, context windows, keeping it fast.
- Retrieval done as math: build honest recall over your notes with lichen-core (free companion).
- Written from a working rig, not a blog roundup. Every claim tested on the machine in front of us.
Pay USDT or ETH on Base to the wallet below, then download: The Local AI Field Manual (PDF). Honor system — the manual explains itself; pay what it was worth to you.
Why it exists
Built because the laptop was too weak for big AI. So we grew one that runs on anything.
That's the whole story. No venture round, no GPU cluster, no API bill — just retrieval done as math, memory modeled as something that grows and fades, and a hard rule: when the mesh doesn't cover a question, it says so, with numbers.
Pay what you want (crypto, no accounts)
The software is free. Payment is how strangers say it was worth something. No accounts, no email, no tracking — just a wallet address.
USDT or ETH on the Base network:
0xe688d2FB4171B80BeaE363451aB13Cf4D625A76F
Base network only — sending on another network loses funds.
Questions
Does it phone home?
No. lichen-core binds to 127.0.0.1 — it is unreachable from the network by construction, and you can verify that yourself with any port scanner. There is no telemetry, no analytics, no account system. If you enable the optional Ollama integration, it talks to the Ollama URL you configure, which is your own local instance by default.
What are the limits?
Single-user by design — no multi-tenant anything. Retrieval is brute-force math over your fact-mesh, which is exactly what makes it honest and offline, but it also means there's a ceiling: very large corpora will get slower, and there's no distributed mode to hide that. We'd rather state the ceiling than pretend it isn't there.
Refunds?
It's free. Download it, run it against your own notes, and pay only after it earns it. There is nothing to refund because nothing is charged up front.
How is this different from ChatGPT or vector RAG?
Two ways. First, the memory is a grown fact-mesh that reinforces with use and decays with neglect — not a static embedding index you re-ingest into. Second, coverage honesty: every answer carries a coverage score, level, and gaps, so "I don't know" is structured output, not a coin flip against a hallucination. And no LLM is needed to remember — retrieval stands on its own math.
Then, if it earns it: add openclaw-factmesh so your agents remember too.