Skip to content

Notes (Memory)

Sidebar → Notes → New

A note repository is a manually managed recall store for agents. Add durable facts in the Web UI, assign the note repository to sessions/channels, and sbot injects relevant notes into the agent context while also exposing a note_search tool for explicit lookup.

Prerequisites

An embedding model is required first: sidebar → Embedding Models → New. See Models.

Configuration

FieldDescription
NameDisplay name for this notebook
EmbeddingEmbedding model used for semantic search (OpenAI, Google, Ollama, Cohere, VoyageAI)

How It Works

  1. Add — You add chunks from the Notes page in the Web UI; long content can be split automatically
  2. Index — Each chunk is indexed with the selected embedding model, or BM25 keyword search when no embedding is configured
  3. Recall — Relevant notes are injected before each turn and can also be searched with note_search
  4. Rank — Results combine semantic/keyword score with time-decay and access-count weighting
  5. Dedup — Near-duplicate inserts are auto-merged when an embedding model is configured

Assignment

Notes can be assigned at multiple scopes (most specific wins):

  • Default per agent — every session/channel using this agent inherits it
  • Per session — override on the chat session
  • Per channel — override per channel

Notes vs Wiki vs Memory

ConceptOwnerBest for
NotesHumans / operatorsFree-form facts the agent recalls semantically or by keyword
WikiHumans / wiki source pluginsStructured knowledge pages with titles + tags
MemoryBackground MemoryLLMDistilled long-term knowledge from past conversations

Released under the MIT License.