Agent Memory FAQs

Question 1

What is Agent Memory?

Accepted Answer

Agent Memory is a local MCP server that provides a shared, durable memory layer for AI agents, storing all facts and contexts in a SQLite database. It allows different AI clients to contribute to and retrieve from a single, persistent memory store.

Question 2

What AI clients can integrate with Agent Memory?

Accepted Answer

Agent Memory is designed for seamless integration with popular AI clients such as Codex and Claude Code. Its setup scripts automatically configure these clients to use the shared memory layer, enabling them to read relevant context and write durable facts.

Question 3

How are user preferences managed within Agent Memory?

Accepted Answer

Agent Memory employs a 'last-write-wins' logic for canonical preferences. When a preference is updated via `memory_upsert`, the latest entry for that specific canonical key and scope becomes the active one, ensuring up-to-date and consistent user preferences across agents.

Question 4

Does Agent Memory support semantic embeddings for memory retrieval?

Accepted Answer

Yes, Agent Memory supports optional semantic embeddings (e.g., via Ollama) to significantly enhance the ranking and relevance of memory retrieval through `memory_search`. It can also operate in a lexical-only mode if embeddings are not configured or desired.

Question 5

How does Agent Memory improve AI agent performance?

Accepted Answer

It enables AI agents like Codex and Claude Code to share persistent context, remember past interactions, and manage canonical preferences. This leads to more consistent, informed, and intelligent behavior across turns and multiple clients working on the same project.

Agent Memory

Agent Memory

Key Features

Use Cases

Key Features

Use Cases