Skip to main content
Recall retrieves supporting clips from your selected history scope and uses a generation model to answer. It does not automatically read every file, application, or uncaptured content. Answers can be wrong.

Prepare Recall

  1. Add a connection and chat preset in Intelligence → Models. In Usage & defaults, select the preset for Recall assistant. You can use Ollama or a supported hosted provider; see model setup.
  2. Optionally enable Meaning Search for semantic retrieval. Recall uses available enabled retrieval sources; embeddings and generation have separate roles.
  3. Select the relevant history scope and filters.
  4. Open Recall from the sidebar, or ask from clipboard search using Recall or its displayed shortcut. Choose a model in the chat header. If no model is configured, open Intelligence to set one up.

Try an answer with evidence

Copy these harmless notes separately:
Ask What time is the Kissa Mori meeting, and how many seats are available? The place name gives exact retrieval a starting point. Open the answer’s numbered citations and check the source text. The goal is a supported answer, not an exact wording or retrieval order. If no evidence appears, confirm the notes are captured and included in the selected scope.

Check and narrow sources

Open citation buttons to inspect excerpts and original clips. Select sources and Use only these clips for a focused follow-up, or deliberately choose Search all history to widen scope. Recall includes initial clipboard evidence when available. Long sources may be shortened or omitted to fit the model’s remaining context. Review partial-source indicators before relying on a reply. Secret and private clips are excluded from Recall evidence. There is no Recall override for that exclusion. A failed privacy check alone does not restrict a clip; see decision rules. Never use real secrets in setup exercises. Citations help inspection; they do not guarantee correct interpretation.

Understand incomplete results

  • No evidence: check scope, filters, captured text, and enabled sources. Try an exact name before a broad question.
  • Incomplete or reduced context: review sources and ask something narrower. Not all retrieved text necessarily fits generation context.
  • Generation unavailable: check the selected connection, credential, model, and feature assignment. For Ollama, also check that the service is running and the model is installed.
  • Incorrect answer: compare sources, refine the question, or evaluate another model. Larger models do not guarantee accuracy.
  • Interrupted answer: reopen the saved chat and retry deliberately. Restarting the app does not replay unfinished tool actions.

Stop and reuse deliberately

Cancel an active answer when needed. Review before copying a reply or applying an action. Supported clipboard changes have an Undo action; available tools depend on the selected model and permissions. Chats and their source snapshots are saved on this device and remain separate from clipboard retention and settings sync. Clearing history does not delete saved chats. Delete a chat to remove its conversation and owned attachments; captured clips remain.

Connect an external assistant

External assistants can retrieve eligible saved clips through the optional local MCP server. They use their own model, while Recall uses the model selected inside ClipsX. Start with read-only access and inspect retrieved sources. Follow Connect assistants with MCP for Codex and Claude Code setup, a first retrieval exercise, permission levels, and token rotation or revocation. See model resources and Troubleshooting.