The brain on disk

Everything wadachi knows lives in one directory you own. No hidden state, no cloud copy. This page is the map.

~/.wadachi/
├── SCHEMA.md      # the wiki conventions — yours to edit
├── index.md       # generated catalog: one line per memory
├── log.md         # append-only history of operations
├── brain.db       # SQLite: metadata, embeddings, beliefs, versions
├── backups/       # automatic backups (migrations, sessions, pre-restore)
├── logs/          # server log — attach to bug reports
├── global/        # cross-project memories (.md)
└── projects/
    └── <name>/    # per-project memories (.md)

Files are the source of truth

Each memory is a plain markdown file with YAML frontmatter. The content of the file is authoritative for what the memory says; brain.db is authoritative for metadata (ids, tags, timestamps, embeddings, beliefs, version history).

---
type: memory
title: Redis pub/sub fan-out choice
project: myapp
tags: ["redis", "sync"]
category: architecture
created: 2026-07-01T10:00:00+00:00
---

Fan-out via Redis pub/sub keeps p95 under budget — see [[#12]].

type is the one field required by the OKF spec (LLM Wiki, Obsidian & OKF); the rest are wadachi's extensions.

You may edit files by hand

It's a feature, not a hack. The parser is tolerant: missing frontmatter, an unclosed ---, tags in JSON or comma format — nothing breaks reading. If a file drifts from canonical form, wadachi doctor --fix regenerates the frontmatter from the database without touching your prose.

The three wiki files

Following the LLM Wiki pattern (LLM Wiki, Obsidian & OKF):

SCHEMA.md — the conventions of your brain. Created once by wadachi init, then never overwritten: edit it, extend it, make it yours. Agents read it.
index.md — the generated catalog: one wikilinked line per memory, grouped by project. Regenerated automatically on every store/delete.
log.md — append-only journal: ## [timestamp] op — detail for every store, delete, merge, obsidian-links run. Grep it to answer "what happened last Tuesday?".

brain.db, briefly

SQLite, one file, versioned schema. Tables: memories, decisions, projects, memory_versions (every prior version of every edited memory), beliefs (the epistemic envelope, Beliefs, staleness & decay), insights, schema_version (migrations, Safety — the whole chain). Embeddings are cached here as blobs so semantic search is instant.

← Connecting your client Memories →