DevFlow

DevFlow MCP provides a persistent, searchable memory for AI agents that support the Model Context Protocol. It stores your project knowledge in a lightweight local database, enabling semantic search, temporal history, and decay-aware confidence for a practical, out-of-the-box memory layer you can rely on in day-to-day development workflows.

How to use

You interact with DevFlow MCP through an MCP client (for example, an LLM with Claude Desktop integration). Start by running the local memory service, then configure your client to connect to it. Use the system to create and retrieve context as you work on features and tasks, and let the memory layer guide your agents with semantically relevant history and decisions.

How to install

Prerequisites: you need Node.js 23+ to take advantage of the latest SQLite features.

Install the MCP CLI globally so you can run the memory server.

# Install globally
npm install -g devflow-mcp

Run the memory server. This creates the database automatically.

dfm mcp

If you want to specify a custom database file location, set an environment variable before starting.

DFM_SQLITE_LOCATION=./my-project-memory.db dfm mcp

Configuration and runtime environment

The system uses a single SQLite database file to store the knowledge graph, with built-in vector search and semantic capabilities. You can customize where the database lives and which embedding model you use by setting environment variables before starting the server.

Key environment variables you may configure:

DFM_SQLITE_LOCATION=./knowledge.db
OPENAI_API_KEY=your-openai-api-key
OPENAI_EMBEDDING_MODEL=text-embedding-3-small
DEBUG=true

Claude Desktop integration and prompts

Connect your Claude Desktop client to the memory server so agents can request context, run semantic searches, and store results. Use workflow-aware prompts that guide agents through planning, task creation, coding, and review while interacting with the knowledge graph.

Prompts are exposed as commands in Claude Desktop (for example, /init-project and /get-context). They return guidance text that tells the AI which tools to call; the actual data operations happen when the AI uses those tools.

Semantic search and history features

Semantic search uses vector embeddings to find related entities by meaning, not just keywords. You can search by query, filter by entity types, and combine semantic results with keyword matching for robust retrieval.

Temporal awareness captures complete version histories for entities and relations. You can retrieve the graph at a specific time, inspect changes, and query past states to understand how decisions evolved.

Confidence decay lets you model how information becomes less certain over time unless reinforced. You can configure half-life and thresholds to keep important data relevant while aging less critical details.

Troubleshooting and diagnostics

If vector search or embeddings fail, the system falls back to text-based search and provides diagnostics to help you identify issues.

Debug mode exposes additional tools to diagnose embeddings and index status. Enable DEBUG to access these diagnostics.

Available tools

create_entities

Create multiple new entities in the knowledge graph with their names, types, and initial observations

add_observations

Append new observations to existing entities to record new facts or updates

delete_entities

Remove entities and their related relations from the graph

delete_observations

Remove specific observations from entities without deleting the entity itself

create_relations

Create multiple relations between entities with enhanced properties such as strength, confidence, and metadata

get_relation

Retrieve a specific relation between two entities and inspect its properties

update_relation

Update an existing relation with new strength, confidence, or metadata values

delete_relations

Remove specific relations from the graph

read_graph

Read the entire knowledge graph structure and contents

search_nodes

Search the graph for nodes based on a query string

open_nodes

Retrieve specific nodes by their names

semantic_search

Search for entities semantically using vector embeddings with optional thresholds and hybrid search

get_entity_embedding

Obtain the vector embedding for a specific entity

get_entity_history

Retrieve the complete version history for an entity

get_relation_history

Retrieve the complete version history for a relation

get_graph_at_time

Get the graph state at a specific timestamp

get_decayed_graph

Get a view of the graph with time-decayed confidence values