Wadachi (轍): the tracks wheels leave in a road. Your AI forgets everything between sessions — re-reading files, re-deriving architecture, re-debating solved problems. Wadachi gives it tracks that stay.
The model re-reads the same files, re-discovers the same patterns — and you type the same reminders, again:
get_context loads project, memories, decisions and what
needs review — ~400 tokens, then it compounds:
Prompt engineering was wording one request well. Context engineering was curating what the model sees before each call. Both run into the same wall — the window fills, quality falls off a cliff, and the usual remedy, summarising to make room, buys that room by throwing precision away. A harness is the scaffolding outside the model that re-initialises the agent step by step: fresh context each step, durable state read back from disk, work resumed exactly where it stopped.
What you learned — it survives the end of a session.
Memories with versions, decisions that keep their rationale and their rejected
alternatives, beliefs that can go stale, a typed graph, sleep. get_context
is the resume step for this layer: one call, and the knowledge is back without
re-deriving it.
What you are doing — it has to survive the end of a context window: the plan for the task in flight, the steps already done, what was tried and failed, where the thread was dropped. Today that state is either lost to compaction or written out by hand as a handover note. Not built yet, and this page says so plainly.
The line that keeps it honest: wadachi never executes anything, and never decides when
something starts. No runner, no sandbox, no scheduler — those belong to whatever
harness drives your agent. reflect, sleep and
consolidate look loop-shaped, but they are background maintenance that
proposes; they never decide that work should begin.
The full explanation →
Most memory tools make knowledge searchable. Wadachi makes it active: it connects, revises, reflects, and learns how you work.
Knowledge stored as plain markdown files with SQLite metadata. Survives across sessions, human-readable, editable by hand. Every edit is versioned — nothing is lost.
Finds memories by meaning, not keywords. Powered by local embeddings via
fastembed — no API calls, no cloud, no cost. Runs entirely on your machine.
The killer tool. get_context detects your project, gathers the relevant
memories and recent decisions, and hands your AI everything it needs in one call.
Not just what you know — what you decided and why. Each decision keeps its rationale and the rejected alternatives, so no choice gets re-debated.
An entity knowledge graph over your brain. recall_associative spreads
activation along citation, semantic and shared-entity edges — surfacing connected
memories even when the words don't match.
Memories are beliefs with an epistemic envelope. Wadachi flags what's gone stale or been superseded, so a session opens knowing what it can no longer trust.
Thinks across memories: surfaces cross-project analogies and non-obvious connections recall can't reach, proposed as insights you accept or reject.
Notices recurring incidents and proposes always-on rules for review — your brain learning not just facts, but how to act. Read-only: it suggests, never edits.
One command wires Wadachi into Claude Code, Claude Desktop, Cursor — any MCP client.
Map filesystem paths to a project name. Wadachi auto-detects which one you're in and scopes memories accordingly.
Every session opens with get_context — the relevant brain is loaded before
the first file is touched.
Bugs, decisions, configs, patterns — saved the moment they're figured out. The brain grows; each session starts faster.
The memory space is crowded and full of good tools. Here is where wadachi actually differs — and where the alternatives are genuinely strong.
| wadachi 轍 | Mem0 / Zep | Supermemory | |
|---|---|---|---|
| Where your memories live | Markdown files on your disk — an Obsidian vault, an OKF bundle | Their cloud (self-host possible) | Their cloud |
| Telemetry | None. Ever. | Product analytics | Product analytics |
| Cost to run | $0 — local embeddings, no API calls | API-metered | Subscription |
| Memory model | Beliefs that go stale, decisions with provenance (why), time-travel (as_of), sleep consolidation |
Facts + graph (Zep: temporal graph, excellent) | Documents + recall |
| Scale & teams | Single-brain, one human — by design | Multi-tenant, production-scale | Team sharing |
| Managed convenience | You run it (one pipx install) |
Hosted, zero ops | Hosted, zero ops |
Honest summary: if you need multi-tenant scale or a managed service, use the alternatives — they are good at it. If you want a brain that is yours, readable on disk, free to run, and that reasons about what it believes — that is what wadachi is for.
Three commands, and you never re-explain your codebase again.
# install (semantic search included, runs locally)
pipx install "wadachi[semantic]"
# guided setup: brain dir, database, Claude Code registration
wadachi init
# restart Claude Code — your AI now remembers