Memory
======

The **Memory** tab controls what the Co-Worker remembers and for whom. You can skip it — without a memory configuration, the Co-Worker does not retain context across sessions.

.. image:: ../images/coworker-memory.png
   :alt: Memory tab with a memory configuration
   :align: center
   :width: 100%

|

.. list-table::
   :header-rows: 1
   :widths: 22 26 52

   * - Setting
     - Options
     - Notes
   * - **Type (backend)**
     - AgentCore, Postgres
     - Locked after the first save. AgentCore is managed long-term memory in AWS; Postgres uses your workspace database and runs extraction on a schedule.
   * - **Strategies** (AgentCore)
     - Semantic, Episodic, User preferences
     - Multi-select. Shown only for the AgentCore backend.
   * - **Strategy** (Postgres)
     - User Level Preference
     - Shown only for the Postgres backend, with an editable extraction prompt and a default 60-minute interval.
   * - **Memory scope**
     - Team memory, Per-member isolated
     - *Team memory* = one shared pool for all users of this Co-Worker. *Per-member isolated* = each person has private context. New Co-Workers default to **Per-member isolated**.
   * - **STM Retrieval Limit**
     - 0 to 1000 (default 10)
     - Maximum short-term events returned at run time. Shown for the Postgres backend only.
   * - **LTM Retrieval Limit**
     - 0 to 1000 (default 10)
     - Maximum long-term records returned at run time. Shown for the Postgres backend only.

.. note::

   Retrieval-limit fields appear only when the backend is **Postgres**; with **AgentCore**, retrieval uses built-in defaults. Memory scope is set here and applies within a single Co-Worker.

Reflection
----------

A **Reflection** toggle — on by default — lets the Co-Worker learn from how people respond to it over time. When it is on, signals such as a user's reaction to one of its messages on Microsoft Teams help it adjust how it responds.

.. tip::

   Reflection works hand in hand with the implicit feedback captured on Teams. For how those reactions are captured and used, see :doc:`observability`.