SQLite long-term memory for Claude, Codex, and other MCP agents

What it does

Memora plugs into MCP-compatible clients like Claude Code and Codex CLI to give agents a SQLite-backed memory layer that outlives the session. It stores memories, TODOs, issues, and structured documents, then retrieves relevant context through semantic search, full-text filters, and a memory_digest tool. A built-in HTTP server renders the growing knowledge graph as an interactive web UI updated in real time via SSE.

The interesting bit

Rather than dumping context into a flat vector database, Memora treats memory as a typed graph with hierarchical sections, auto-linked cross-references, and fragment-level document integrity. It can run entirely offline with local embeddings or sync encrypted SQLite to Cloudflare D1 and S3-compatible stores for access across machines.

Key highlights

• Pluggable embeddings: OpenAI, sentence-transformers, or TF-IDF, switchable at startup.
• Hierarchical organization with typed edges, importance boosting, and cluster detection.
• Document fragments (claims, plans, risks) are individually searchable yet guarded from accidental deletion.
• Live graph server includes a timeline panel, action history, and a RAG chat panel for streaming queries against stored memories.
• Optional cloud sync via Cloudflare D1 or S3/R2 with client-side encryption and compression.

Caveats

• Changing the embedding model after memories exist requires manually rebuilding vectors and cross-references.
• Local embedding mode depends on roughly 2GB of PyTorch dependencies.
• Event notifications for inter-agent communication are poll-based, not push-based.

Verdict

Worth exploring if you want Claude or Codex to remember project context across sessions. If you are not using an MCP-compatible agent client, or you only need ephemeral prompt stuffing, this will be overkill.