Memory & knowledge

Two systems handle long-term context. Memory holds facts about you. The knowledge base holds facts about the world that the assistant should be able to look up.

How knowledge enters the brain

You never sit down to write a memory. Every conversation quietly seeds the brain, and you stay in control of what lands.

  1. Stage 01

    Conversation

    Web, Telegram, Slack. Every channel feeds the same pipeline. Voice notes are transcribed on arrival.

  2. Stage 02

    Extract

    A background pass distils stable facts from each turn. What you said matters, not the chat verbatim. Mentions of yourself become identity candidates; mentions of others become entity candidates.

  3. Stage 03

    Consolidate

    Light pass dedups near-duplicates against existing memories and the KB. Deep pass synthesises narratives, prunes stale rows, and adjusts confidence.

  4. Stage 04

    Land

    Each row lands with tags, source, and a pointer to the episode that produced it. Preferences surface first; situational context fills in around them.

Memory

Memory is per (user, assistant). The assistant extracts patterns from what you say ("we're based in Hong Kong," "we invoice net-30," "the Q3 launch ships Sept 18") and stores them. On every turn, a curated summary of relevant memories is included in the system prompt.

How a memory is shaped

01Preference

stable facts about how you work ("prefers markdown", "invoices net-30"). Surfaced first when retrieval is relevant.

02Context

situational facts the assistant has observed ("Q3 launch ships Sept 18", "Acme renewal closes this Friday"). Most rows live here.

03Provenance

every row points back to the episode that produced it (the chat turn, voice memo, or connector event), so you can trace any belief to its source.

Your controls

Every memory is editable. Settings → Privacy → Memories shows all stored memories. Search, edit, or delete one-by-one or wholesale. Asking the assistant "forget that" works too.

How the brain answers

Every turn fans in identity, memories, knowledge, tools, and the live session, then assembles one prompt the model can actually use.

Your message
  • 01Identity
  • 02Relevant memories
  • 03Knowledge base
  • 04Tools & connectors
  • 05Recent session
Context
Selected per-turn, ranked by relevance.
03

Compose

System prompt + selected context + tool catalogue + your turn.

04

Model

Standard / Pro / Max tier per turn. Set in the chat header. Background work (extraction, embedding, classifiers) always runs Standard.

05

Reply

Streamed back to your channel. The loop begins again at Extract.

Knowledge base

The KB is per-workspace, not per-user. Use it for facts the assistant should know (proposal voting rules, product specs, FAQ answers). The assistant has searchKnowledge / browseKnowledge / readKnowledgeEntry / addKnowledgeEntry tools to retrieve and append. Connect a GitHub repo from Studio → Knowledge to sync markdown docs automatically.

Sensitivity tiers

public
internal
confidential

Each KB entry is tagged public, internal, or confidential. The assistant's clearance gates what it can read across every channel, including the public API. To expose only public KB to API consumers, set the assistant's clearance to public; for an internal integration, use an assistant with the matching clearance.