agentic-box/memora

pip install git+https://github.com/agentic-box/memora.git
pip install "memora[local]" @ git+https://github.com/agentic-box/memora.git   # offline local embeddings (~2GB PyTorch)

Usage

• •memora-server # stdio MCP mode; auto-runs when configured in Claude Code
• •Add to .mcp.json (Claude Code) or ~/.codex/config.toml (Codex) with env like MEMORA_DB_PATH / MEMORA_STORAGE_URI / MEMORA_GRAPH_PORT
• •memora-server --transport streamable-http --host 127.0.0.1 --port 8080 # HTTP alternative
• •Core pattern: memory_absorb agent work, then memory_digest(topic) to retrieve relevant memories, TODOs/issues, edges, and source IDs

Key features

• ✓Semantic search with three embedding backends (openai default, sentence-transformers offline, tfidf); hybrid full-text/date/tag queries; LLM deduplication with merge strategies
• ✓Knowledge graph with interactive live visualization server (port 8765; Mermaid rendering, timeline, history panels) and a cloud-hosted option on Cloudflare Pages with D1
• ✓Chat panel with RAG and tool calling (search/create/update/delete memories by chat); event notifications; memory insights with stale detection
• ✓Document storage as searchable fragment trees (claims, plan items, references, risks) with integrity guards against accidental modification
• ✓Typed memory linking (references/implements/supersedes/extends/contradicts/related_to), cluster detection, export/import with merge, Neovim plugin, cloud encryption (MEMORA_CLOUD_ENCRYPT)

Best for

Giving Claude Code or Codex agents durable, graph-structured cross-session memory with a visual knowledge-graph UI.

Caveats

• ⚠OpenAI embeddings / LLM dedup / chat require OPENAI_API_KEY (or an OpenAI-compatible base URL); local embeddings are an extra ~2GB install
• ⚠Rebuild embeddings manually (memory_rebuild_embeddings + memory_rebuild_crossrefs) after changing the embedding model
• ⚠Default chat model is deepseek/deepseek-chat; chat model must support function calling