Your agents forget everything. Memori treats that as a bug.

What it does

Memori sits between your agents and their existing infrastructure to persist structured memory across sessions. It wraps standard LLM clients—OpenAI, Anthropic, Gemini, and others—to automatically capture conversation and execution context in the background. The system organizes memories around entities, processes, and sessions, then recalls relevant state when needed without bloating prompts.

The interesting bit

The project pitches itself as infrastructure rather than a framework: it is datastore-agnostic and offers a bring-your-own-database option, so you do not have to migrate off your existing data infrastructure to use it. It also claims to outperform Zep, LangMem, and Mem0 on the LoCoMo long-conversation benchmark while using about 4.97% of the token footprint of full-context prompting.

Key highlights

• LLM-agnostic wrappers for Python and TypeScript that intercept and persist chat completions automatically
• Drop-in plugins for OpenClaw and Hermes agents, plus an MCP server for Claude Code, Cursor, and similar tools
• Structured memory augmentation in the background across entities, processes, and sessions (facts, preferences, relationships, rules, skills)
• Deployable across managed cloud, single-tenant cloud, VPC, and on-premises
• LoCoMo benchmark result: 81.95% accuracy with ~1,294 tokens per query, which the project claims is about 4.97% of a full-context approach

Caveats

• The README pushes Memori Cloud heavily—most quickstarts assume an API key, and the free tier is rate-limited
• BYODB is mentioned but self-hosting details live in external docs, not the README
• Quota limits for the free tier and Advanced Augmentation are mentioned, but exact caps are not listed

Verdict

Worth evaluating if you need cross-session agent memory that works across multiple LLMs and frameworks. Keep in mind that the README emphasizes the hosted cloud path; BYODB and on-premises deployment are mentioned but not detailed inside the repository itself.