Getting started
Local single-user deployment, free edition. From an empty directory to first search in about five minutes.
Prerequisites
- Python 3.12 and a writable local directory (default ~/allmymind/amm/data).
- Models are only needed for digging and grouping; capture and search work without one.
- To connect a host you need an MCP client: Claude Desktop, DSH, or any compatible implementation.
Install
Install into a dedicated venv from source, then initialise. amm init creates the database if absent.
Add this to your host MCP config and restart the host.
Configuration
A few environment variables and one .env. Usually only the data directory and model endpoint matter.
| Variable | Meaning | Default |
|---|---|---|
AMM_DATA | Data directory (SQLite / chroma / backups) | ~/allmymind/amm/data |
AMM_LLM_BASE_URL | Model endpoint | https://api.deepseek.com/v1 |
AMM_LLM_API_KEY | Model key | (empty) |
PYTHONPATH | Required in the MCP env block | .:<venv>/site-packages |
MCP tools
21 tools and 4 resources. Writes are dry-run by default (confirm=false) and only commit after confirmation.
| Tool | Purpose | Arguments |
|---|---|---|
leaf_add | Store a leaf in the pending layer | text / node_id |
search_semantic | Search by meaning; returns leaves with their path | query / top_k |
tree_get | Fetch a tree four levels deep | root_id |
memory_suggest | Offer a card only: count, length, last stored | query / budget |
memory_commit | Return the text after the user chooses | topics |
hole_put | Write a hollow (trunk only, never searched) | trunk_id / text |
The daily three
Drop a couple of leaves, search once in a while to check it finds the right thing, group when they pile up. Do not expect it to get smart by itself — a forest is grown, not built.
- Capture: amm add "a sentence" — it lands in the pending layer.
- Find: you do not need to know which tree; just say the sentence.
- Group: amm plan proposes, amm apply commits; for folders use amm dig.
- Press “save” weekly to back the whole library up to data/backups/.
FAQ
Where is memory stored? Is it uploaded?
Entirely local: four SQLite tables plus a local Chroma index. No cloud required, unless you configure a cloud model.
Can several agents share one forest?
Point them at the same AMM_DATA. Note that SQLite is single-process — do not write from multiple instances at once.
What if it remembers something wrong?
Find the leaf and archive it, or cut the tree / prune the branch. Only “purge” is irreversible, and it shows the count first.
Why 409 DUPLICATE when digging?
That file was imported before. Fingerprints live in earth; the default policy rejects, and you can switch to confirm.
No module named 'torch'?
PYTHONPATH is missing the vector library's site-packages. The MCP env block must set PYTHONPATH explicitly.