Skip to main content
A memory is a unit of knowledge stored in Velixar. It can be a fact, a preference, a decision, a conversation summary — anything your AI should remember.

​
Memory structure

{
  "id": "mem_abc123",
  "content": "User prefers dark mode and metric units",
  "user_id": "user_123",
  "tier": 0,
  "type": "preference",
  "tags": ["settings", "ui"],
  "metadata": {},
  "score": 0.92,
  "created_at": "2026-02-16T10:30:00Z",
  "updated_at": "2026-02-16T10:30:00Z"
}
FieldTypeDescription
idstringUnique memory ID (UUID)
contentstringThe memory text — this is what gets embedded and searched
user_idstringOwner of this memory, for multi-user isolation
tierintegerStorage tier (0–3), controls priority and retention
typestringfact, preference, context, or custom
tagsarrayFilterable labels
metadataobjectArbitrary key-value data
scorefloatRelevance score (only in search results)

​
Memory types

  • fact — Objective information: “User’s company is Acme Corp”
  • preference — User choices: “Prefers concise responses”
  • context — Situational: “Currently working on Q1 budget”
  • custom — Any string you define
When you store a memory, Velixar generates a vector embedding of the content field. Search queries are also embedded, and results are ranked by cosine similarity. This means you search by meaning, not keywords:
  • Query: “what color does the user like?” → matches “User’s favorite color is blue”
  • Query: “display settings” → matches “User prefers dark mode”