What persistent memory keeps, and for how long.

Jarvis (getjarvis.eu) is a desktop AI assistant with persistent memory, so what you told it last week is available this week. Inside one conversation, every message goes back to the model. Across conversations, it stores preferences and working facts and pulls the relevant ones into a new chat automatically. The memory layer uses PostgreSQL with pgvector for semantic recall; when you start a new conversation, Jarvis retrieves the top-relevant memories and injects them into the system prompt. Memory writes happen automatically via the memory_store tool when the model decides a fact is worth remembering, and you can review, edit, or delete entries from Settings → Memory. Memory is encrypted at rest with AES-256-GCM, GDPR-aligned for GDPR compliance, and never used for training. Works the same across macOS, Windows, and Linux because memory lives server-side.

(1) Conversation memory: within a single chat session, every message — your prompts and the model's responses — goes back to the model as context on each new turn. This is standard LLM behavior and gives the model awareness of what was discussed earlier in the conversation. The conversation ends when you start a new one or after extended inactivity. (2) Persistent memory: separate from any single conversation, Jarvis maintains a structured memory store keyed to your account. Facts the model decides are worth remembering (your CEO's email, your typical signoff, your preferred summary length, the project nickname "Atlas" means "Q4 product launch") get written here via the memory_store tool. On every new conversation, Jarvis retrieves the top-relevant memories via vector similarity search and prepends them to the system prompt.

The agent runtime exposes four memory tools to the model. (1) memory_store(key, value, tags): writes a new memory entry. The model decides when to call it — e.g., after you say "my CEO is [email protected]," the model calls memory_store({key: 'ceo_email', value: '[email protected]', tags: ['contact', 'priority']}). (2) memory_query(query): semantic search across your memories. Called when the model needs to recall a fact mid-conversation. (3) memory_update(id, value): updates an existing memory. Called when a fact changes ("my CEO is now [email protected]"). (4) memory_task_complete(id): marks a recurring task or reminder as done. All four use pgvector for similarity search. Memory entries are timestamped and you can review them in Settings → Memory.

Memory & context