# Memories

Canonical page: `https://nanthai.tech/features/memories`

Tier: Pro

AI that remembers your preferences, projects, and context across every conversation.

Edge can save facts about you — your writing style, current projects, relationships, and goals. Memory is categorised, searchable, and fully transparent.

## How it works

1. **AI extracts facts as you chat** — After every AI response, Edge analyses the exchange and extracts up to 4 atomic facts worth remembering — preferences, project context, writing style, relationships, goals, and more. Facts are deduplicated and quality-filtered before being offered for saving.
2. **You choose what to keep** — In "Ask First" mode, new facts queue as pending memories for you to review. Approve, edit, or discard each one individually — or bulk-approve the entire queue. In automatic mode, facts are saved immediately. You can disable extraction entirely.
3. **Memories are injected into every conversation** — Up to 10 always-on memories are injected into every chat. Up to 12 additional memories are retrieved via semantic search — finding facts most relevant to your current conversation. Together they appear in three sections of the system prompt.
4. **Memories evolve over time** — When new information reinforces an existing memory, the confidence score goes up. When it contradicts one, the old memory is superseded and the new one takes its place. Temporary memories expire automatically — work context after 90 days, transient facts after 21 days.

## What you can do

- **10 memory categories** — Every memory is filed under one of 10 categories: identity, writing style, work, goals, background, relationships, preferences, tools, skills, and logistics. Browse and filter by category in the Memories view.
- **3 retrieval modes** — Choose how each memory is used: Always On (injected into every chat, up to 10), Contextual (retrieved via semantic search when relevant — the default), or Disabled (stored but never surfaced).
- **Import from documents** — Upload PDFs, Word documents, Markdown, or plain text files. Edge extracts memory candidates from the document, then lets you review and commit each one before they're saved.
- **Pending review queue** — In "Ask First" mode, all new extractions queue as pending memories. Edit any field, approve individually, or bulk approve/reject the entire queue. Nothing is committed until you confirm.
- **Configurable extraction model** — Choose which AI model runs the memory extraction step. Use any model available in your account — useful if you want a cheaper model for extraction and a more powerful one for chat.
- **Pin for priority** — Pin any memory to boost its retrieval priority. Pinned memories get a +0.35 boost in the ranking algorithm and are more likely to be included in the contextual batch.
- **Per-persona scoping** — Memories can be scoped to all personas or only specific ones. A memory about your coding style can be restricted to your developer persona and excluded from your creative writing persona.
- **Automatic expiry and consolidation** — Work context memories expire after 90 days. Transient facts expire after 21 days. A daily consolidation job at 3:00 UTC cleans up expired memories and merges reinforced duplicates.
- **Reinforcement and supersession** — When new information matches an existing memory, confidence scores increase. When it conflicts, the old memory expires and links to the new one — the full chain is preserved for transparency.
- **Semantic search retrieval** — Contextual memories are retrieved using vector embeddings (OpenAI text-embedding-3-small). The system finds the most relevant facts for your current conversation, not just keyword matches.
- **3 system prompt sections** — Retrieved memories are injected as three named sections: Response Preferences (how you like answers formatted), User Profile (who you are), and Relevant Context (what's relevant right now).
- **Fully transparent** — Every memory is visible, editable, and deletable. See when it was saved, which conversation it came from, how many times it's been used, and its confidence score.
- **Full control** — Edit or delete any memory at any time. Clear all memories with one action. Turn off extraction, turn off injection, or turn off memories entirely — each is a separate setting.
- **Pro-only with data protection** — Memories are a Pro feature. If your Pro subscription lapses, all memories and their embeddings are purged automatically. No data lingers after downgrade.

## Perfect for

- Never re-explaining your tech stack — the AI remembers your languages, frameworks, and conventions and tailors suggestions accordingly.
- Maintaining writing consistency — the AI remembers your preferred tone, style, and vocabulary across every writing task, scoped to your writing persona.
- Project context that persists — your AI knows you're working on a specific project, its goals, and its current status without you having to re-brief it each session.
- Importing your CV — uploading a resume as a document, reviewing the extracted memory candidates, and letting the AI know your background once for all future conversations.
- Ask First workflow — reviewing each week's accumulated memory queue in one session: approving the useful facts, editing the ones that need tweaking, and discarding the rest.
- Persona-scoped knowledge — keeping your work memories in your professional persona and your personal preferences in a separate personal persona, with no cross-contamination.

## Related features

- Personas (`/features/personas`)
- Knowledge Base (`/features/knowledge-base`)
- Chat Defaults & Favorites (`/features/chat-defaults`)