Memory System

The memories Table

Memory is stored in a simple key-value table scoped to each team:

memories (
  id        serial PRIMARY KEY,
  teamId    text NOT NULL REFERENCES teams(id),
  key       text NOT NULL,
  content   text NOT NULL,
  createdAt timestamp,
  updatedAt timestamp
)

One row per key per team. Think of it as a persistent scratchpad that agents can read and write across all conversations.

The memory Tool

The memory tool (in lib/ai/tools/shared/memory-tool.ts) exposes four commands:

The mode field on update: "overwrite" (default) replaces the content, "append" adds to it.

The Special core Key

The core memory key is the team's base context — brand voice, business description, writing style, client preferences, recurring topics.

Agents that use prepareCall fetch this key and inject it into their system prompt:

--- TEAM CORE MEMORY ---
We are a B2B SaaS company targeting CTOs. Our tone is direct, technical, and jargon-free.
We post on LinkedIn twice a week. Never use generic AI buzzwords.
------------------------

This context persists across every conversation without the user having to re-explain their brand each time.

To update core memory, the user can ask any agent: "Update our core memory to note that we're launching a new product next month" — the agent calls memory({ command: "update", key: "core", content: "..." }).

Searching Memory

The search command uses a SQL ILIKE query with a LIMIT 3:

SELECT * FROM memories
WHERE teamId = $1 AND content ILIKE '%query%'
LIMIT 3