How to Configure Persistent Memory with MemUbot for Long-Running Agents

What you need

A server with Docker or a Python environment (3.10+).

MemUbot installed: `pip install memubot` or using the Docker image.

A model provider API key.

Understanding of what context your agents need to remember.

Step 1: Understand the three memory layers

Short-term memory holds current session context — recent messages, current task details. It resets when the session ends.

Working memory stores active task information that persists across sessions but has a configurable expiration.

Long-term memory captures learned preferences, approved knowledge, and persistent patterns. It is indexed and searchable.

Step 2: Configure retention policies

Each memory layer has configurable retention: max items, TTL (time to live), and importance scoring.

Short-term: keep 50 most recent items, TTL of 1 hour.

Working memory: keep 200 items, TTL of 7 days.

Long-term: unlimited items, no TTL, but importance score below 0.3 is auto-archived.

Step 3: Set up the memory backend

MemUbot uses SQLite by default for local deployments. All memory is stored on your infrastructure.

For production, configure the Postgres backend for better scalability and query performance.

The memory vault is compatible with Obsidian — you can browse and edit the agent's memory directly.

Step 4: Configure data sensitivity rules

Sensitive data can be restricted to short-term memory only — it will not persist in working or long-term storage.

Define patterns: 'Any data matching email addresses, API keys, or phone numbers stays in short-term only.'

Audit rules ensure compliance with data privacy requirements like GDPR right to erasure.

Step 5: Test context retrieval

Start a session, give the agent several instructions with specific preferences.

End the session and start a new one. Ask the agent about the preferences from the previous session.

Verify it correctly retrieves long-term memories and recognizes the user.

Step 6: Optimize for your use case

For research agents: extend long-term retention and add automatic importance scoring based on citation frequency.

For customer support: shorten short-term TTL and increase working memory capacity for active ticket context.

For personal assistants: enable cross-session learning so the agent adapts to user preferences over time.

FAQ

Configure MemUbot Memory

1. Install MemUbot: pip install memubot
2. Configure three memory layers with retention policies
3. Set up the memory backend (SQLite or Postgres)
4. Define data sensitivity rules
5. Test context retrieval across sessions
6. Optimize retention settings for your use case