Memora

Memora is an open-source MCP (Model Context Protocol) server that provides AI agents with persistent, collective memory across sessions. Built in Python and released under the MIT license, it stores structured facts in SQLite with optional cloud sync to Cloudflare D1, S3, or R2, and exposes 44 MCP tools for creating, searching, linking, and digesting memories. The project reached v0.5.5 in September 2026 and has accumulated over 725 stars on GitHub.

What It Is

Memora sits between AI agents (such as Claude Code or Codex CLI) and a durable storage backend, acting as a memory layer that agents can read from and write to via the Model Context Protocol. Rather than each agent session starting from scratch, Memora lets agents absorb facts from prior work, retrieve relevant context through semantic search, and maintain a living knowledge graph of what they know. It is listed in the Awesome Claude Code collection, signaling community recognition as a useful tool in the Claude ecosystem.

How the Absorb-Digest Workflow Works

The core workflow centers on two operations:

• memory_absorb — Feed raw facts into the store. An LLM classifies each incoming fact against existing memories as a duplicate, update, contradiction, related item, or new entry. Duplicates are skipped; updates supersede old knowledge without deleting it (supersession lineage); contradictions and related items are linked. A dry_run mode previews changes before committing.
• memory_digest(topic) — Retrieve a bundled context package for a topic: relevant memories, open TODOs and issues, related graph edges, and source IDs — all in one call, reducing the number of round trips an agent needs to orient itself.

This absorb-then-digest pattern is designed for multi-agent workflows where a leader agent coordinates workers, each contributing facts that accumulate into a shared store.

Storage and Deployment Architecture

Memora supports two primary deployment paths:

• pip / stdio — Install memora-mcp from PyPI and spawn it as a stdio child process from .mcp.json. Local SQLite is the default backend; cloud sync to S3/R2 or Cloudflare D1 is configured via environment variables.
• Container / HTTP service — Run Memora as a detached HTTP service using Apple's container CLI (requires Apple Silicon and macOS 26). A bundled proxy script (memora_proxy.py) holds a stable local address in front of the container's reassigned IP, and a macOS LaunchAgent supervises the proxy.

Multi-database routing (MEMORA_DATABASES) lets a single process serve multiple workspaces, each reaching its own store at /mcp/<name>. Tool profiles (MEMORA_TOOL_PROFILE) expose subsets of the 44 tools — full (all 44), leader (19), or agent (12) — so worker agents are not burdened with destructive maintenance tools.

Search, Embeddings, and Intelligence

Semantic search is powered by one of three embedding backends: OpenAI (default, included), sentence-transformers (offline, installed via pip install memora-mcp[local]), or TF-IDF (keyword-based, always available). Embeddings and the LLM used for deduplication and chat are configured separately, allowing different providers for each role. The README explicitly warns that OpenRouter has no embeddings endpoint and that mixing it into the embedding path causes silent TF-IDF fallback; MEMORA_EMBEDDING_STRICT=1 converts that silent degradation into a hard failure.

Additional intelligence features include:

• LLM deduplication — Find and merge duplicate memories with AI-powered comparison returning a verdict, confidence score, and suggested action.
• Memory insights — Activity summaries, stale TODO/issue detection, consolidation candidates, and LLM-powered pattern analysis.
• Memory linking — Typed edges (implements, supersedes, contradicts, etc.), importance boosting, and cluster detection.

Knowledge Graph and Visualization

A built-in HTTP server (default port 8765) serves an interactive knowledge graph with a details panel, timeline panel, history panel, and a RAG-powered chat panel. The chat panel supports streaming LLM responses with tool calling, so agents or users can create, update, and delete memories directly from the UI. For Cloudflare D1 deployments, a hosted graph on Cloudflare Pages is available with WebSocket real-time updates and Cloudflare Zero Trust access control, eliminating the need for SSH tunneling.

Update: v0.5.5 — Local SQLite Primary with D1 Replication

The latest release, v0.5.5, published September 24, 2026, introduces local SQLite as the primary store with D1 replication. Earlier releases addressed D1-specific performance issues such as vector scan page size limits and absorb failures caused by Cloudflare's 30-second per-request ceiling. The project's direction continues toward multi-agent, multi-database deployments with fine-grained tool exposure and robust cloud-sync options.