Skip to content

Memory ​

GhostAI Memory allows the AI to remember your preferences, facts, and instructions across conversations. All memories are privacy-protected with PII scanning, tokenization, and encryption.

What is Memory? ​

Memory is GhostAI's ability to learn from your conversations and remember important context. When you mention a preference ("I use Python 3.12"), state a fact ("Our API uses REST"), or give an instruction ("Always use formal tone"), GhostAI can remember it and apply it in future conversations.

Enabling Memory ​

  1. Go to Settings > Configuration > Memory
  2. Toggle Enable Memory on
  3. Choose which categories to remember (Preferences, Facts, Instructions, Context)
  4. Optionally enable Auto-Extract to have GhostAI learn from conversations automatically

Memory can also be enabled during onboarding for paid plans.

How Memories Are Created ​

Automatic Extraction ​

When auto-extract is enabled, GhostAI analyzes your conversations after each response and identifies information worth remembering. Extraction runs as a background task using the same LLM provider configured for chat, with a dedicated system prompt optimized for JSON memory extraction.

Manual "Remember This" ​

Click the bookmark icon on any message to save it as a memory. You can also add memories directly from the Memory page.

Organization Knowledge ​

Org admins can seed a knowledge base that all members have access to. This is ideal for company standards, coding conventions, and domain terminology.

SaaS Integrations ​

Connect your SaaS tools (Rippling, Slack, Jira, Salesforce) in Settings → SaaS Integrations to automatically populate the AI's memory with your organization's data. Employee directories, team structures, and organizational context are synced on a schedule and bridged into memory with PII filtering.

Connector-sourced memories include source attribution — the AI cites where its knowledge came from (e.g., "[Source: Rippling, synced 2h ago]").

See SaaS Integrations for setup and configuration.

Managing Your Memories ​

Access the Memory page from the Quick Access section at the top of the sidebar (brain icon). You can:

  • View all your memories with category badges and source information
  • Search by keyword or semantic similarity
  • Filter by category, source, or pinned status
  • Edit any memory inline
  • Pin important memories so they're never auto-expired
  • Tag memories for organization
  • Delete individual memories or bulk-delete all (GDPR)
  • Export all memories as JSON

Conversation Memory ​

Each conversation automatically extracts key decisions, entities, and context. This helps overcome the message window limitation. Conversation memories are ephemeral by default but can be promoted to your personal memory.

Toggle memory off for a specific chat using the brain icon in the chat header ("off the record" mode).

Organization Knowledge Base ​

For Business and Enterprise plans, org admins can:

  • Manually add knowledge items
  • Bulk import via CSV or JSON
  • Mark items as "mandatory" (always injected) or "contextual" (semantic match only)
  • Review auto-extraction suggestions

Groups and Tags ​

Groups (Team plan and above): Formal teams/departments with shared knowledge bases. Create groups from the Memory page or Admin Dashboard.

Tags: Lightweight collections for organizing memories.

Auto-Tagging ​

When auto-extract is enabled, GhostAI automatically generates 1-3 keyword tags for each extracted memory (e.g. #python, #coding-style, #healthcare). Tags are inferred from conversation context by the extraction LLM and stored alongside the memory.

Managing Tags ​

From the Tags tab in the Memory Manager, you can:

  • Create tags manually with the tag name input
  • Delete tags you created (removes the tag from all associated memories)
  • Toggle sharing -- mark a tag as "Shared" to make its tagged memories visible to your team, or "Private" to keep them personal

Tagging Memories ​

  • When creating a memory: Use the tag picker below the content field. Type to search existing tags or create new ones inline.
  • On existing memories: Each memory card shows its #tag badges. Hover to reveal an x button to remove a tag, or click the + button to assign additional tags from a dropdown.

Filtering by Tag ​

Use the tag filter dropdown in the filter bar to view only memories with a specific tag. Combine with category filter and pinned-only for precise views.

Vector Explorer ​

The Vector Explorer provides an interactive visualization of your memory embeddings, showing how memories relate to each other in semantic space.

Accessing the Explorer ​

  • Sidebar: Click the compass icon in the Quick Access section
  • Memory Manager: Switch to the "Explorer" tab
  • Admin Console: Platform admins can access the "Embeddings" tab for a cross-org view

Projection Methods ​

  • PCA (default): Fast, stable layout that preserves global structure. Ideal for initial exploration.
  • t-SNE: Click "Refine" for better cluster separation. Slower but reveals tighter groupings.

Features ​

  • Scatter plot: Each point is a memory projected from 384-dimensional embeddings to 2D
  • Color coding: Points colored by scope (user, tenant, group, etc.) or category (preference, fact, instruction, context)
  • Hover tooltips: Quick preview of memory content with category and confidence
  • Detail panel: Click any point to see full content, metadata, tags, and nearest neighbors with similarity scores
  • Search-to-highlight: Type to highlight matching memories in the plot
  • Filters: Filter by scope, category, or search term
  • Force graph overlay: Toggle "Connections" to see similarity edges between related memories. Adjust the threshold slider to control how many connections are shown.
  • Multi-source view: Toggle between Memories (circles), File Chunks (diamonds), and Content Library (squares) to visualize all your embeddings together
  • Quick actions: Pin/unpin and delete memories directly from the detail panel

Admin Mode ​

Platform admins see all organization memories in a single view from the Admin Console's Embeddings tab. An amber banner indicates admin mode. This is useful for understanding organization-wide memory coverage and identifying knowledge gaps.

Memory and Privacy ​

  • All memories pass through PII detection and tokenization
  • Embeddings use a local model (never sent externally)
  • Advanced retrieval architecture: At scale, memory recall uses HNSW indexing for sub-linear approximate nearest-neighbor search, two-level routing embeddings (compressed routing vectors plus full semantic embeddings) to narrow and rank candidates, and Memory Interleave — optional iterative recall_memory tool rounds during chat — for multi-hop reasoning when enabled.
  • Dynamic token budgeting: Injected memory text scales with the active model’s context window — by default 15% of context, capped at 8,000 tokens (with a small floor so short contexts still get usable recall).
  • Hybrid search: Keyword (BM25) and semantic retrieval are fused with RRF (Reciprocal Rank Fusion) so both exact phrases and paraphrases surface reliably.
  • Recall sensitivity: You can tune how aggressively memories match your conversations — Conservative, Balanced, or Broad — via a user-configurable similarity threshold (see Settings or the Memory API).
  • Tiered caching: A Redis hot tier accelerates repeated routing and recall paths; the database is the warm source of truth; long-term or infrequently touched material follows your retention policy as a cold archive layer.
  • Shared embedding model: Memory embeddings and RAG document embeddings both use the same all-MiniLM-L6-v2 model (384 dimensions), ensuring consistent semantic representation across personal memory and file retrieval. The RAG system extends this with hybrid BM25+semantic search (Reciprocal Rank Fusion), GPU-accelerated batch encoding (CUDA/MPS auto-detect), and HNSW approximate nearest neighbor for files with >100K chunks. See File Intelligence for details.
  • Raw content is encrypted at rest
  • Memories respect your organization's data retention policy
  • Delete all memories at any time via the Memory page or API

Memory Limits by Plan ​

PlanMemory LimitGroupsOrg Knowledge
FreeNoneNoneNone
Personal500NoneNone
Team2,000/userUp to 5None
Business10,000/userUp to 505,000 items
EnterpriseUnlimitedUnlimitedUnlimited

Memory does not consume credits.

FAQ ​

Q: Does memory send my data externally? A: Memory embeddings use a local model and never leave your machine. Memory extraction uses the same LLM provider you've configured for chat. If you use Ollama (local), nothing leaves your network. If you use a cloud provider like OpenRouter or Gemini, extraction prompts are sent to that provider — the same one already handling your chat messages.

Q: Can my admin see my personal memories? A: No. Personal memories are visible only to you. Admins can see org knowledge base items and usage statistics.

Q: What happens when I hit my memory limit? A: New memories won't be stored, and auto-extraction pauses. You'll see an upgrade prompt.

Q: Can I use memory via the API? A: Yes. The Memory API is available on Team plans and above. See the API Reference.

SOC 2 Type II Certified | BAA Available | HIPAA-Ready