Mesh-Context-Layer

Development teams struggle with "agent amnesia" where LLMs lose context between sessions or repeat mistakes other agents already solved. Mesh Context Layer provides a high-speed, persistent semantic memory layer that allows agents to share facts, decisions, and instructions across different tools and sessions.

What it does

• Persistent Storage of atomic facts, architectural decisions, and workflows using a SQLite-backed hybrid memory system.
• Hybrid Search combines vector embeddings, BM25 text search, and exact token matching to find relevant context.
• Conflict Management identifies contradictory information during storage and provides tools to resolve data inconsistencies.
• Privacy Controls enable marking sensitive data as local-only to redact it from exports and logs.
• Lifecycle Management uses TTL decay to automatically lower the confidence of stale information over time.

How it works

1. Initialize the environment using mesh setup to configure namespaces and local embedding models.
2. Store discoveries using mesh_learn to save atomic facts with specific types like decision or instruction.
3. Recall relevant context at the start of a task using mesh_recall to query the shared memory pool.
4. Update existing knowledge using mesh_update to refine facts without losing historical access data.

Frameworks & tools

This skill interfaces with Python 3.10+, SQLite FTS5, and ONNX. It integrates directly with MCP-compatible clients including Claude Desktop, Cursor, and Gemini CLI.

Why this beats prompting it yourself

Manual prompting lacks persistence across different agent instances and cannot scale to thousands of project facts. This skill automates the vectorization, deduplication, and decay logic required to maintain a reliable long-term memory.

Use cases

• Cross-Agent Handover where a coding agent stores a bug fix and a DevOps agent later recalls the deployment steps.
• Architecture Log to track why specific design choices were made to prevent regressive changes.
• Shared Config Management for storing staging URLs, port mappings, and environment-specific flags.

Known limitations

Requires Python 3.10 or higher. Performance depends on SQLite FTS5 support and local disk I/O for the vector database.