Skip to main content

Memory

Nimbus maintains a per-repository memory of coding conventions, architectural patterns, and task outcomes. Memory is used during planning to give Claude more context about your codebase.

List entries​

nimbus memory

Shows the last 10 memory entries for the current repository:

  memory  github.com/acme/api

  [a3f2c1]  2026-04-25  All API handlers use async/await. Never use sync DB calls.
  [b7d4e2]  2026-04-24  Auth middleware is in api/middleware/auth.py. Always use the
                         @require_auth decorator, never check request.user directly.
  [c9f1a3]  2026-04-23  Tests use pytest-asyncio. Fixtures are in conftest.py.
  ...

Delete an entry​

nimbus memory --delete a3f2c1

Use the ID shown in square brackets.

Query memory for a specific repo​

nimbus memory --repo ~/other-project

What gets stored automatically​

After every completed task, Nimbus calls Claude to extract and store:

  • Coding conventions observed or reinforced
  • Architectural patterns used
  • Libraries and frameworks encountered
  • Whether the task succeeded or failed (and why, if failed)

Each entry is at most 150 words. Nimbus stores it alongside a timestamp and the task description that triggered it.

How memory affects planning​

When you run a task, Nimbus queries memory for the 8 most relevant entries using semantic search and injects them into the planning prompt. This means Claude Opus has context like "always use the @require_auth decorator" before it generates a plan — without you having to repeat yourself.

Storage location​

Memory is stored in ChromaDB at ~/.nimbus/chroma/, in a per-repo collection named repo_memory_{repo_id}. It's separate from the code index and persists indefinitely until you delete entries.