Your AI agent’s memory finally gets a git log

What it does

Memoria is a persistent memory layer that sits between AI agents and a database, tracking every memory change with Git-like semantics—snapshots, branches, merges, and rollbacks. It stores preferences, facts, and episodic context across conversations, and retrieves them through a hybrid of vector and full-text search. The project ships with agent-specific steering rules that tell tools like Cursor, Claude Code, or any MCP-compatible client when to store, correct, or purge memories.

The interesting bit

The backing trick is MatrixOne’s native Copy-on-Write engine, which makes zero-copy branching and point-in-time rollback practical without loading entire datasets into memory. Memoria also auto-detects contradictions and quarantines low-confidence memories, treating data integrity as a first-class concern rather than an afterthought.

Key highlights

• Git-level version control native to the storage layer: branch an experimental context, merge it back, or revert if the agent drifts off-course.
• Hybrid semantic retrieval combining vector search with full-text search.
• Self-governance features that flag contradictions and isolate questionable memories.
• Local embedding model support for fully offline operation.
• MCP-native protocol support; works with Kiro, Cursor, Claude Code, Codex, and Gemini CLI.
• Backed by an arXiv research paper on applying Git workflows to cloud-native database storage.

Caveats

• The repo description claims it reduces hallucinations, but the README only spells out contradiction detection and quarantine as the mechanism; the direct link to fewer hallucinations is assumed, not demonstrated.
• It is unclear whether the cloud and self-hosted tiers expose identical versioning capabilities.

Verdict

Worth a look if you are building agentic tools that need durable, auditable context across long sessions. Skip it if you are looking for a simple key-value cache—this is a database-backed versioning system with a CLI and steering rules, not a drop-in Redis replacement.