devmind-mcp

Stores and retrieves AI assistant context, conversation history, and working memory in a local SQLite database that persists across multiple MCP tool invocations and client sessions. Uses a schema-based storage model where context entries are indexed by conversation ID, timestamp, and content type, enabling the assistant to maintain coherent state without relying on in-memory buffers or external cloud services. Implements automatic garbage collection and context windowing to prevent unbounded database growth.

Exposes a registry of available MCP tools and provides a standardized interface for other MCP tools to discover, invoke, and chain tool calls with automatic context passing. Implements a schema-based tool discovery mechanism where each registered tool declares its input/output types, and the orchestrator validates arguments before invocation, catching type mismatches and missing required parameters. Supports both synchronous and asynchronous tool execution with error handling and result caching.

Enables context and memory state to be shared between different MCP clients (e.g., Claude Desktop, custom agents, IDE plugins) by exposing context as queryable MCP resources that any connected client can read and write. Implements a simple versioning scheme where each context update increments a version number, allowing clients to detect stale data and request fresh context. Uses MCP's resource subscription mechanism to push context updates to interested clients in real-time.

Retrieves conversation history from the SQLite context store with support for filtering by conversation ID, time range, message type, and content keywords. Implements pagination to handle large conversation histories without loading entire datasets into memory. Returns results as structured JSON with metadata (timestamps, sender, message type) alongside content, enabling downstream processing and analysis.

Captures and stores detailed traces of agent decision-making processes, including intermediate reasoning steps, tool selections, and outcome evaluations. Each trace entry includes the agent's input, reasoning chain, selected action, and result, enabling post-hoc analysis of agent behavior. Implements a hierarchical trace structure where multi-step agent workflows can be represented as nested traces, with each level capturing the reasoning at that abstraction level.

Automatically manages the size of context windows by summarizing older conversation segments and compressing them into condensed representations. Implements a sliding window approach where recent messages are kept in full detail while older messages are progressively summarized. Uses configurable thresholds to determine when summarization triggers, balancing context freshness with token efficiency.

Supports multiple independent conversations within a single devmind-mcp instance by using conversation IDs as namespace keys. Each conversation maintains its own context, history, and traces, with no cross-contamination between conversations. Implements query filters that automatically scope all context operations to the specified conversation ID, preventing accidental data leakage.

Provides a mechanism for agents to extract and store structured data (facts, decisions, extracted entities) from unstructured conversation content. Implements a schema-based storage model where extracted data is validated against declared schemas before storage. Supports querying extracted data by type, enabling agents to retrieve previously extracted facts without re-processing conversation history.