Use personal memory and understand source and analysis coverage.
Reviewed · 2026-10-12
On this page · 5
Manage personal research memory
Use Personal AI memory to add, edit or forget preferences, decisions, hypotheses, research lessons and task-specific conditions. Project or grant scope remains private to you; it does not publish team knowledge. Current instructions override old preferences. Source-backed memories are not reused after their source changes, is deleted or becomes inaccessible.
Recall uses relevant excerpts from completed personal consultations and external discussions actually delivered through supported tools. Personal consultations, new research strategies, code consultations and daily plans/reviews share this service. Shared conversations, public reports and translation do not receive wholesale private history. Use the existing reviewed report sharing copy and recipient selection for team knowledge. Code knowledge retains the repository commit, source reports and verification status.
Excluding a conversation stops implicit recall elsewhere but preserves original history and saved outputs. Re-inclusion never overrides a previous reset cutoff. Memory reset and source deletion are separate. Previous outputs remain history, but forgotten memories cannot become future evidence.
Registration, search and exclusions use no extra AI calls. Context tokens are included in normal response usage. Selected memory text is capped at 8000 characters; this is not exhaustive recall. Source data, indexes, settings and exclusions belong in encrypted external recovery backups. Apply the latest exclusion ledger before reopening a restored application. Backup and provider retention periods are distinct from immediately stopping reuse.
1. Recall across conversation windows
- Open Profile menu → Settings → Personal AI memory. Use personal memory across conversations is enabled by default.
- Ask a question and wait for a completed answer. Your question and answer are stored for your account. In another window, mention the earlier topic or distinctive terms, for example: “Use the validation sequence we discussed to plan the next experiment.”
- Existing completed consultations are indexed gradually. Unindexed history is still searchable and older matches have no fixed date cutoff.
- Open references to check which conversations and materials informed the answer. This retrieves related history; it does not retrain the model or read every conversation in full on every request. Repeat critical requirements in your question.
- Automatic memory uses your own completed consultations. Other users’ private consultations and support tickets are excluded. Deleted conversations and currently inaccessible evidence are not reused. AI suggestions are not confirmed research facts.
2. Pause or reset memory
- Clear the checkbox and select Save memory settings to pause recall, including implicit history within the same window. Re-enabling recalls history after the most recent reset.
- Select Reset personal memory and confirm to delete saved memory and establish a new cutoff. Answers already in progress cannot restore pre-reset memory.
- Original consultation history stays available for viewing and export, but pre-reset history is not automatically reindexed. New consultations establish new memory.
- To delete the original conversation, use its history-management controls. Reset cannot recall previously downloaded or externally archived copies. Personal memory is excluded from shared lab Drive archives.
3. Make sources available for consultation
- Register and authorize GitHub, Notion, Drive and other connections on the grant/project, then check collection status. Consultation searches currently readable collected materials, GitHub Projects work, reports, documents, events, channel/direct messages, notices, discussions and grant/project descriptions.
- To collect an entire Drive account, select All Drive accessible to this Google account and set the lookback period to 0. Select a folder for narrower coverage. Scope participants may see collected materials, so consider personal files before selecting the whole account.
- Save, authorize Google and select Collect now. Large listings continue over multiple runs. Inspect extracted content, source dates and omissions. An uncollected linked file has not been read by AI.
- Include document names, scope, dates and key terms, or insert a [[ citation. All allowed sources are candidates, but a response receives selected excerpts with at most 80 evidence items and a byte limit. It does not read every file in full at once.
4. Automatic text and visual analysis, with limits
- Edit the Drive connection and select Automatic analysis (default) under PDF, image and embedded document image analysis. Existing connections are eligible unless explicitly disabled. On request only stops automatic queue enrollment.
- Text: TXT, MD, CSV, JSON and related text, DOCX, PPTX slides and notes, XLSX sheets and cached values, HWP 5 body text, HWPX, PDF text layers, ODT, ODS and ODP. Google Docs, Sheets and Slides are exported to Office formats. Macros and formulas never execute.
- Visual analysis processes PDF pages, PNG/JPEG/WebP/TIFF/BMP files and embedded raster images in Office, HWPX and ODF packages. The connection owner needs AI permission, lab AI enablement and daily allowance. Exhausted allowance retries the next day. Previously collected files are enrolled gradually.
- Open File visual analysis on a material to inspect progress and notes. A unit is a PDF page or an embedded Office image. This does not fully render Office layouts, vector diagrams/SmartArt, HWP images or transcribe audio/video. Verify numbers and plots against originals.
- Limits include 20 MB per file, 200,000 extracted characters, 300 PDF/image units, and pixel/compression limits. Google export size limits also apply. Encrypted/distribution HWP, HWP 3, legacy DOC/XLS/PPT and unsupported files may be unreadable. Export a readable PDF, DOCX or HWPX and reconnect it.
- Analysis notes are distinguished from originals in consultation retrieval. Pending, failed or partial analysis is not complete coverage. Recheck collection and analysis after source or permission changes.