THE COMPANY BRAIN FOR CODING AGENTS

Your agent's memory remembers facts.
It doesn't track the work.

Built-in agent memory and CLAUDE.md are notes — they reload whole, go stale, and have no idea what you're working on: your tasks, your projects, what's in progress, what depends on what.

Marvis is your Company Brain: a queryable memory plus the tasks, projects, and decisions that make up your work — all linked, so your agent knows the state of things, not just a pile of notes.

Install · uv tool install marvisx-cli Or: tell your agent to set it up

WHAT A NOTES FILE CAN'T DO

Tracks the work, not just facts

Tasks, projects, milestones — what's in progress, blocked, done. Your notes file has no idea.

Retrieved, not reloaded

Your agent asks for the one decision it needs — not the whole stale file dumped into context.

Self-consolidating

Contradictions get superseded over time, not silently stacked on top of each other.

Related & queryable

"What depends on this — what breaks if I change it?" has a deterministic answer. Prose doesn't.

SAME MEMORY, BUT IT KNOWS THE STATE OF THINGS

Your agent's native memory

## decisions.md
- use Postgres for storage
- actually SQLite, single writer
- auth via JWT cookie
- moved auth to header tokens
- naming: kebab-case slugs
- ...
[ + 380 more lines ]

> the agent reloads all of it —
or misses the line that mattered.

Marvis Company Brain

agent asks: "what storage did we
settle on — and what breaks if I change it?"

→ Decision: SQLite, single-writer
  supersedes "Postgres" · 2025-04
→ Linked: task #128 "bound reindex
  memory" · project acme-api
→ Blast radius: 3 modules call
  db.write()

> one exact answer + everything
  attached to it.

● Local-first ● Free, open-core (BSL 1.1 → Apache 2.0) ● Claude Code · Codex · Cursor, via MCP