ContextGraph

Overview

ContextGraph provides a durable memory backend for coding agents and multi-agent teams, featuring governed shared memory, context compilation, and reactive delta compaction. It enables agents to survive context pressure without losing state through structured checkpoints and maintains visible state in repository-local directories. The server implements the MCP protocol to expose these memory management capabilities.

Try asking AI

After installing, here are 5 things you can ask your AI assistant:

• contextgraph_compile_context

  Compile a governed, token-budgeted context pack from mixed agent memory

  Try:Compile context for payment service issues with 1000 tokensTry:Create a context pack about project status for me
• contextgraph_store

  Store a claim in agent memory with provenance and access controls

  Try:Store that payment service is migrating from REST to gRPC for Q2Try:Remember that we had an incident with connection pool exhaustion
• contextgraph_recall

  Retrieve memories matching a query with relevance scoring

  Try:Recall information about payment service latency issuesTry:What do we know about the API migration project?
• contextgraph_register_agent

  Register a new agent with the memory system

  Try:Register a new agent named 'ops-bot' for operations teamTry:Create an agent called 'research-assistant' for the research department
• contextgraph_create_session

  Create a new session for an agent to record events

  Try:Start a session for the payments refactor projectTry:Create a new coding session for API migration
• contextgraph_record_session_event

  Record an event (decision, constraint, task, etc.) in a session

  Try:Record a decision to keep the REST API stableTry:Note that we need to add migration tests
• contextgraph_checkpoint_session

  Create a checkpoint from a session to save its current state

  Try:Checkpoint the current session before making major changesTry:Save the current state of the coding session
• contextgraph_fork_session

  Create a new session that inherits from an existing checkpoint

  Try:Fork the current session to experiment with gRPC implementationTry:Create a branch from the last checkpoint to try a different approach
• contextgraph_sync_memory_directory

  Sync session state to a local .contextgraph/ directory

  Try:Sync the current session state to my workspaceTry:Update the memory directory with the latest changes
• contextgraph_get_context_pack

  Retrieve a previously compiled context pack

  Try:Get the context pack I compiled earlier about payment issuesTry:Retrieve the project status context pack
• contextgraph_explain_context_pack

  Get explanations for why claims were included or excluded in a context pack

  Try:Explain why some claims were excluded from the payment service packTry:Show me why claims conflicted in the last compilation

Comparable tools

Installation

pip install contextgraph

{
  "mcpServers": {
    "contextgraph": {
      "command": "python",
      "args": ["-m", "contextgraph.server"],
      "env": {}
    }
  }
}

FAQ

What makes ContextGraph different from vector memory?

ContextGraph provides structured memory with provenance, freshness, access control, and explainable retrieval rather than just raw text storage.

How does reactive delta compaction work?

Instead of collapsing sessions into summaries, it records structured events as checkpoints and compiles delta packs that preserve decisions, tasks, and state across context boundaries.

Compare ContextGraph with