Model server
Daily Recall sends every request to one OpenAI-compatible chat completions endpoint. Screenshot analysis needs a model that reads images. Summaries, digests and search answers are text only.
Settings › Model
- Base URL including
/v1, for examplehttp://localhost:8080/v1. Requests go to<base URL>/chat/completions. A missing/v1shows up as HTTP 404. - Model: the name the server expects. If you leave it empty, the name the server reports is stored.
- API key: only if the server requires one. It is sent as
Authorization: Bearer <key>. - Test Connection sends a small picture (a red circle and a blue square) and shows the reply and the response time. If the server can’t read images, it shows the error.
llama.cpp
Start llama-server with a vision model and its projector file:
llama-server -m model.gguf --mmproj mmproj.gguf --host 127.0.0.1 --port 8080 -c 131072
Check it from the shell:
curl -s http://localhost:8080/v1/chat/completions -H 'Content-Type: application/json' -d '{
"messages": [{"role": "user", "content": "Reply with the single word: ready"}],
"max_tokens": 20, "chat_template_kwargs": {"enable_thinking": false}}' | jq -r '.choices[0].message.content'
By default llama-server answers one request at a time, so a search question waits behind a running screenshot analysis or summary. Start it with -np 2 to answer two in parallel. Each slot gets half of the context set with -c, so double -c as well.
LM Studio
LM Studio works with the same settings. Use the server address LM Studio shows, with /v1, and load a vision model.
Thinking
Prompts 2 to 6 have a “Let the model think” switch. Screenshot analyses never think.
| Thinking on | Thinking off | |
|---|---|---|
| Sent | enable_thinking: true | enable_thinking: false, reasoning_effort: "none" |
| Summary output limit | 64,000 tokens | 3,000 tokens |
| Marked moment output limit | 8,000 tokens | 800 tokens |
llama.cpp follows enable_thinking. LM Studio ignores it and needs reasoning_effort, so both are sent. A <think>…</think> block in a reply is removed before it is stored.
Context size
A summary sends the whole day in one request. A full day at the 5-minute interval holds about 100 screenshot descriptions and a few hundred window spans. With thinking on, the reply can take up to 64,000 tokens on top of the prompt, so give the server a large context, for example -c 131072.
If the server rejects a summary for its context size, Daily Recall retries once with a shorter log. If that fails too, the overview shows the error. Raise the server’s context and regenerate.
A summary cut off by the token limit is stored as a version marked incomplete, with a warning in the overview.
When the server is down
Screenshots wait in a queue that survives restarts. While the server is unreachable they stay pending and are retried, starting at 30 s and doubling up to 15 minutes. The menubar menu shows how many are pending or failed, with “Retry failed”.