FAQ — the honest answers
Privacy & trust
Does wadachi send my data anywhere? No. No telemetry, no analytics, no cloud calls, no "anonymous usage statistics". Embeddings are computed locally. The only network access in the product's life is pip downloading the package. This is design rule #1, not a toggle.
How do you improve it without telemetry, then? People tell us. There's a "Share your setup" issue template — voluntary feedback is the only channel, which is exactly the trade we want.
Can I audit what it stores?
cat your brain. It's markdown in a folder (The brain on disk). log.md lists every
operation ever performed.
What if I want out?
pipx uninstall wadachi and your brain remains as plain files — readable forever,
openable in Obsidian, importable anywhere as an OKF bundle. Exit is as easy as
entry.
Costs & performance
What does it cost to run? $0. Local embeddings, no API calls. The optional
entity graph uses your local claude CLI (your existing plan, not metered API).
How big can a brain get? The design target is a personal brain: hundreds to a few thousand memories. Search is instant at that scale; the graph builds in well under a second at ~100 nodes.
Does it work offline? Entirely, after the one-time model download.
Behaviour
Will it ever change my memories on its own? No. Bookkeeping (index, log, access counters) happens on your own writes; everything cognitive — merging, retiring, insights — is proposed and waits for you (get_context — start here, every session, Sleep — consolidation, proposed).
What happens on upgrades? Migrations with automatic backups; you can snapshot
first with wadachi export. Full chain in Safety — the whole chain.
Semantic search in my language? The default model is English-optimized but handles Italian and other European languages reasonably. Keyword mode is language-neutral. A multilingual model option is on the roadmap.
Can several agents share one brain? Several clients on one machine, yes — they're all the same server. Several machines: one writer at a time (Connecting your client).
Comparisons
vs Mem0 / Zep? They're good, production-scale, managed services with multi-tenant APIs. Wadachi is single-brain, local, files-you-own, with a cognitive model (beliefs, provenance, time-travel, sleep) they don't aim for. If you need a hosted memory API for your SaaS, use them. If you want your brain on your disk that reasons about what it believes — that's wadachi.
vs Obsidian + plugins? Obsidian is a wonderful editor for humans. Wadachi is a memory server for agents that happens to speak Obsidian's format (LLM Wiki, Obsidian & OKF). They compose: same folder, two doors.
vs just using CLAUDE.md? CLAUDE.md is instructions — a few hundred lines the
model always reads. Wadachi is knowledge — thousands of memories retrieved by
relevance. You need both; wadachi's get_context is how the right slice of the
second reaches the model without drowning it.
Project
Who's behind this? One engineering student and his AI pair programmer. The whole history is in the repo's CHANGELOG — including the day it was renamed from Engram after discovering four name collisions (ask why() about it 😉).
When is v1.0? When the tool API and schema have been stable for a month. The clock is running.
How do I contribute? PRs welcome (CONTRIBUTING.md has the four rules and the dev setup). Not a coder? Feedback is the most valuable contribution there is.