claude-memory

The Claude memory tool, and the consolidation loop that keeps memory clean. One repo, one backend, two layers, and every run makes a real API call.

• The primitive. Declare the GA memory tool (memory_20250818) and service the commands the model emits (view, create, str_replace, insert, delete, rename) against a real, path-checked local store.
• The pattern. A two-day dreaming loop. On day one, task agents write memory as they work. A consolidation pass verifies each claim against the transcripts, drops the unsupported ones, organizes what is left, and cross-links it. Day two runs the same cohort on the consolidated memory and measures the deltas.

Both layers go through the same backend in claude_memory/memory.py.

Run it

pip install -r requirements.txt
ANTHROPIC_API_KEY=... python run.py              # the dreaming loop
ANTHROPIC_API_KEY=... python run.py memory_tool  # the memory tool primitive

Every run calls the real Anthropic API. Without a key it fails fast with a clear error and a non-zero exit. There is no offline mode.

Layout

Path	What it holds
claude_memory/memory.py	the one backend: the path-checked store and the six commands
claude_memory/tool.py	the primitive demo: declare the tool, service the commands
claude_memory/loop.py	the two-day loop
claude_memory/dreaming.py	the consolidation pass (verify, organize, cross-link)
claude_memory/agents.py	the task agents that write memory on day one
evals/eval_dreaming.py	the gate: day two must beat day one, and the pass must drop at least one unsupported claim

Note on scope

Anthropic's Dreaming is a research preview, not a public API. This implements the dreaming pattern on GA primitives (the memory tool, context editing, route-by-consequence model selection), so it runs today, and it says so. It claims only what it runs.

License

MIT. See LICENSE.