Skip to content
Capabilities

Add a capability manually

Wrap a memory library behind cap.memory/v1: scaffold, try, connect, validate, prove.

With Nullius installed or running from source and a model provider; know the library's storage and recall operations.

Create the files

Scaffold my-memory in the user catalogue
nullius standard new capability my-memory
nullius_catalogue=$(nullius paths --json | python3 -c 'import json, sys; print(json.load(sys.stdin)["catalogue"])')

The new folder is capabilities/my-memory/ below that catalogue. It survives application upgrades. No package resources or repository tests need changing for a personal addition.

capability.json · excerpt of the scaffold
{
  "kind": "memory",
  "start": {"command": ["python3", "{here}/capability.py"]},
  "offers": [],
  "proven": {}
}

Try the protocol

Two requests to the scaffold
python3 "$nullius_catalogue/capabilities/my-memory/capability.py" <<'JSONL'
{"op":"add","session":{"id":"s1","text":"The project codename is Cedar."}}
{"op":"recall","question":"What is the project codename?"}
JSONL
Reply
The scaffold {} then {"memory": ""}
Implemented {} then {"memory": "The project codename is Cedar."}

One JSON request per line on stdin, one reply per line on stdout, logs only on stderr.

Connect the library

Request The wrapper Reply
add Stores session.text, or text built from session.turns {}
recall Searches question, formats the records as text {"memory": "..."}

Examples: app/nullius/catalogue/capabilities/mem0/capability.py calls an upstream memory library; app/nullius/catalogue/capabilities/keyword-memory/capability.py is the protocol alone, BM25 over sessions.

Declare dependencies

Source Fields
Python packages Pinned versions in install.pip
A repository upstream.repository, the full commit in upstream.tested, install.project in the checkout
Model calls A relay model configuration
capability.json · Mem0's Python dependencies (excerpt)
{
  "install": {
    "python": "3.12",
    "pip": ["mem0ai==2.2.1", "fastembed==0.9.1", "spacy==3.8.18"]
  }
}

Validate and prove

  1. Validate (ok, nothing run):

    Validate the definition
    nullius standard check \
  "$nullius_catalogue/capabilities/my-memory/capability.json"

  2. Conformance:

    Run conformance
    nullius standard conform my-memory \
  --provider openai --model gpt-5-mini

  3. A small benchmark:

    Two units of LongMemEval
    nullius standard fit longmemeval
nullius standard run longmemeval my-memory \
  --provider openai --model gpt-5-mini --limit 2
Rule
The empty scaffold Fails the recall checks
Success Passing checks write offers and proven, never by hand
Scale Inspect the stored and recalled text before raising the limit

To contribute the entry to the bundled catalogue, copy its whole folder into app/nullius/catalogue/capabilities/ in a source checkout, run the relevant suites and refresh python3 app/tests/catalogue_docs.py --write. The generated docs read only bundled definitions.

Conformance and a benchmark make paid model calls

Each check and unit calls the model through the relay.