---
source_hash: 6ba8d18f55a76066896261067b2b112e0775c2bd4a156ae07c64856b48bdd4bc
post: /posts/ari-os/advanced
shape: doctrinal
---

# Context Pack: Advanced ARI-OS capabilities

Agent reference for the rewritten guide. Only quoted excerpts are verbatim. The old blanket default-off claim was corrected using public source commit 85bba5fccf42549b0507bc852ac8aaf47844b812. These are inspected code paths, not successful backend executions.

## Verbatim excerpts

> The core brain doesn't need every optional capability switched on.

> It does not store that text in memory.

> It doesn't call a vision model.

## Defaults and command namespaces

The checked defaults are KG off, EARS off, LENS off, wander on and predictive on. Prefetch shares the predictive gate. Saved configuration and environment overrides can change those defaults; inspect the installed state.

Operations use `python3 -m ari_os.tools.cortex`. Settings use `python3 -m ari_os.tools.arios cortex ...`. The nine original Code blocks remain byte-for-byte in the article; additional prose corrects what their commands actually do.

## Six capabilities

- Knowledge graph: `kg extract` requests extraction subject to the KG toggle and backend requirements. `kg list` reads entities; `kg stats` reads entity, relation and chunk-link counts. The two readers are not gated like extraction.
- Prediction: bare `predict` reports that no action was selected. Choose `--cluster`, `--signals` or `--cwd`. It respects the predictive gate.
- Prefetch: `prefetch --cwd <dir>` retrieves context for the named directory and uses the same gate. Inspect the returned material rather than assuming it is useful.
- Wander: `wander --focus "<text>"` looks for an associative tangent. Control it with `python3 -m ari_os.tools.arios cortex wander on|off`. Its checked default is on.
- EARS: `ears <file|url>` transcribes local audio or retrieves a supported YouTube transcript, then prints the text. It checks EARS and enabled Cortex LLM configuration. It does not save a memory; review the output and use a separate memory write path if desired. Plain transcription does not need a vision model, but its own backend still must be available. Enable with `python3 -m ari_os.tools.arios cortex ears on`.
- LENS: `lens <slug>` and `brain.lens(slug)` retrieve a local card or indexed content after the LENS gate. There is no model call. `brain.see(image_path, mode='visual')` additionally checks enabled Cortex LLM configuration and needs a working vision backend. Missing cards and unavailable models yield messages, not a successful capability. Enable LENS with `python3 -m ari_os.tools.arios cortex lens on`.

## Embeddings and limits

`python3 -m ari_os.tools.arios cortex embeddings auto|google|ollama|off` selects the preference. These values save a control-panel preference. Inspected retrieval/indexing paths instantiate the Ollama client directly; this inspection does not establish that they honour auto, a named provider or off. Storage locality does not establish processing locality.

## Evidence and references

- [Configuration](https://github.com/PUSHINGSQUARES/ARI-OS/blob/85bba5fccf42549b0507bc852ac8aaf47844b812/ari_os/tools/cortex/config.py).
- [CLI](https://github.com/PUSHINGSQUARES/ARI-OS/blob/85bba5fccf42549b0507bc852ac8aaf47844b812/ari_os/tools/cortex/cortex.py).
- [MCP tools](https://github.com/PUSHINGSQUARES/ARI-OS/blob/85bba5fccf42549b0507bc852ac8aaf47844b812/ari_os/tools/cortex/mcp_tools.py).
- [Core brain guide](/documents/ari-os/brain).

SurfaceList and FAQ were corrected with the prose so they no longer repeat the old default/LENS conflict. The diagram, PatternDetect and Pullquote remain.