Memory

Pandora’s memory is configured on the Memory page in the web UI. After changing memory settings, restart Pandora for the changes to take effect.

How It Works

When memory is enabled, Pandora uses two complementary systems:

• Working memory — a persistent scratchpad of important facts about you and your context. Available immediately from the first message in any thread, so your agent always has baseline context.
• Observational memory — background agents watch your conversations and build up observations (concise notes about what happened, what you discussed, and what matters). These observations persist across all threads.

As conversations grow, older message history is compressed into observations. When observations themselves grow large enough, they’re further condensed through reflection. The result is a multi-tier system:

1. Working memory — key facts, available instantly in every thread
2. Recent messages — exact conversation history for the current exchange
3. Observations — compressed notes from past conversations
4. Reflections — condensed observations when memory grows very long

This happens in the background — you don’t need to do anything. Pandora decides what to remember automatically.

Enabling

Toggle memory on the Memory page. Memory is enabled by default.

Model Override

By default, the background memory processes use the same model as your chat model. You can optionally specify a different model on the Memory page using the provider/model format (e.g. google/gemini-2.5-flash).

A fast model with a large context window works best for memory processing, since it runs in the background and handles potentially large amounts of conversation history.

Monitoring

The Memory page includes a Status card that shows memory activity:

• Observation capacity — shows how full the current observations are. When the bar fills, existing observations are condensed automatically.
• Activity indicator — shows whether Pandora is currently processing memories or idle.
• Stats — how much has been processed overall, how many times memories have been condensed, and when the last processing happened.

Status updates automatically every few seconds.

Working Memory

The Memory page shows the current working memory — key facts your agent has noted. Pandora manages this automatically, but you can also edit it directly to correct or add information.

Viewing Observations

The Memory page shows the current observations that Pandora has built up. These are read-only — Pandora manages them automatically. You can see what it remembers about you and your conversations.