agent-memory

Persistent long-term memory and memory-lifecycle services for AI assistants.

The repo has two main layers:

• Hippocampus is the storage and retrieval service. It exposes a small HTTP API over Mem0, with local SQLite/in-memory fallbacks.
• Memory Governor is the policy layer in front of it. It decides what should become durable memory, tracks recall and outcome signals, consolidates working observations, and drives the dreaming/reflection lifecycle.

Older docs and service names may still use Sacred Brain; that is the historical deployment name, while agent-memory is the repository and subsystem boundary.

Features

• Store and semantically retrieve durable memories with provenance and metadata.
• Observe high-volume agent activity without turning the memory store into a transcript archive.
• Promote, consolidate, rerank, prune, and protect memories using recall, relevance, recency, query diversity, confidence, and outcome signals.
• Keep short-lived working observations separate from durable memory.
• Run nightly dreaming/reflection over recent activity and recalled memories.
• Fall back to local SQLite or in-memory storage when the configured Mem0 backend is unavailable.
• Expose CLI, HTTP, MCP, hook, bot, and file/org integration surfaces without making any one agent harness the source of truth.
• Include pytest coverage for the Hippocampus API, Governor, dreaming, recall statistics, scopes, routing, and related lifecycle behaviour.