agentic-box/memora

What it solves

Memora provides AI agents with a persistent, collective memory layer. It prevents agents from losing context across sessions by allowing them to store, organize, and retrieve structured facts, TODOs, and documents in a durable way.

How it works

It operates as a Model Context Protocol (MCP) server that manages a knowledge graph. It uses SQLite for local storage with options for cloud synchronization (S3, R2, D1). The system uses an "Absorb" process where an LLM classifies incoming information to avoid duplicates and link related facts. Retrieval is handled through semantic search (vector embeddings), full-text queries, and a specific memory_digest function that bundles relevant memories and open issues for a given topic.

Who it’s for

Developers building AI agents that require long-term memory, cross-session context, and the ability to maintain a structured knowledge base of project-specific facts and tasks.

Highlights

• Persistent Graph Memory: Stores information as a graph with typed edges and hierarchical organization.
• Supersession Lineage: Updates knowledge by superseding old facts rather than deleting them, maintaining a history of changes.
• Semantic Search: Supports hybrid search using vector embeddings (OpenAI, sentence-transformers) and full-text filtering.
• Absorb & Deduplication: Uses LLMs to automatically classify, merge, and deduplicate new information being fed into the memory.
• Document Storage: Handles markdown documents as searchable fragment trees (claims, risks, plans).
• Visualization: Includes a built-in HTTP server for interactive knowledge graph visualization.