Build Your Portrait

The life portrait gathers your identity, goals, circumstances, preferences and plans from your chat history, organized into domains such as work, family and health and into topics within them. Every fact traces back to the original chat. This route needs no GPU, since extraction runs on the API model you configure.

First complete installation, the model backend and data preparation.

Run

From the project root:

weclone-cli build-profile

Choose the purpose when asked, or pass it directly:

• --purpose assistant: for a personal assistant. Extracts the portrait and events.
• --purpose clone: for a clone. Also extracts your speaking style and needs the speaking style dependencies.

Before it starts, the command line shows the result directory, what will be extracted and the model backend, and asks you to confirm. After you confirm, chat content is sent to that model service.

Option	Description
--purpose assistant|clone	The purpose. Asked interactively when omitted
--output-dir <dir>	Result directory, which must be new. Defaults to a new timestamped directory under dataset/res_csv/agent/profile_runs/
--max-context-tokens <limit>	Input token limit for memory organization and the hierarchy. The Codex backend reads the model's context length by default. The API backend requires it

Stages

Stage	What happens
1. Prepare chat data	Text messages are grouped by contact and cut into conversations, written to people/
2. Extract memories	Profile memories and event memories are extracted from each conversation, written to distill/
3. Organize memories	Domains are induced, memories are classified, grouped into topics and summarized, and every fact gets a privacy level and a time scope, written to memory_organization/
4. Build the hierarchy	The portrait shown and reviewed in the web UI, written to profile_hierarchy/
5. Domain summaries	One summary per domain, see Domain Summaries
6. Speaking style	--purpose clone only, written to style/, see Speaking Style

While chat data is prepared, one sender's messages less than two minutes apart become one message, and a silence of more than ten minutes starts a new conversation. Very short conversations are skipped to save model calls.

What the portrait holds

Each fact belongs to a dimension and carries:

Field	Description
Dimension	Demographics and identity, needs motives and goals, resources and constraints, key plans and commitments, interests and preferences
Domain and topic	Induced by the model from your chats, such as work, family and health
Privacy level	Public, normal, private or sensitive. Decides what assistants can be given
Time scope	Long-term, current, outdated, or an event at a given time
Sources	The conversations backing the fact, viewable in the web UI
Confidence and importance	Averages of the source scores

No fact is dropped for privacy during extraction. Every fact is kept with a privacy level, and you decide who gets what when you review and connect assistants.

Cache and reruns

Model answers are cached by input content in dataset/res_csv/agent/llm_cache.sqlite. If a run fails partway, rerun it into a new result directory and only the missing parts are requested. Delete the file to request everything again.

Contact numbers are kept in dataset/res_csv/agent/contacts.json, which only grows. A contact keeps its number across new exports and reruns.

Result directory

dataset/res_csv/agent/profile_runs/<time>/
├── people/                # chat snapshot
├── distill/               # extracted profile and event memories
├── memory_organization/   # memories with domains, classes and summaries
├── profile_hierarchy/     # the portrait and your review in the web UI
└── style/                 # speaking style (--purpose clone only)

In encrypted mode these files end in .enc. When the run is done, open the web UI with weclone-cli server to review the portrait.

Running a single stage

For debugging or catching up, run a stage on its own:

weclone-cli organize-memories --help                    # domains, classes, topics and summaries
weclone-cli distill-style --input-dir <result-dir>/people  # speaking style

Request settings

Setting	Description
codex_exec_args.model, effort	Model and reasoning effort for the Codex backend
codex_exec_args.batch_size	Concurrent requests for the Codex backend
codex_exec_args.request_interval_seconds, capacity_cooldown_seconds	Interval between requests, and the wait when the model is busy
agent_distill_args.batch_size, timeout, max_tokens	Concurrency, timeout and output limit for the API backend
agent_distill_args.max_samples_per_window, max_content_chars	Short conversations are packed into one request, up to this many conversations and this many characters of content. Longer conversations go in a request of their own

Every request to a model service is logged in logs/llm_audit/, see Security.