Gnosis: A knowledge base capturing context and decisions beyond code implementation

The primary friction in AI-assisted development isn't the code generation; it's the context decay. Every time you start a new session, the agent lacks the history of why certain paths were rejected or what hidden constraints exist in the codebase. Gnosis tackles this by transforming the AI agent from a stateless coder into a historian. By forcing a 'plan-act-review' loop, it ensures that institutional knowledge is captured in real-time rather than relying on the whims of a developer to write a post-mortem document that will inevitably go stale. Technically, the approach is refreshingly pragmatic. Instead of a heavy database or a proprietary cloud silo, Gnosis uses JSONL files stored directly in the repository. This is a smart move for version control; the 'memory' of the project evolves alongside the code via git commits. The use of SQLite's FTS5 for full-text search provides the necessary performance for retrieval without adding operational complexity. It’s a local-first architecture that respects the developer's existing workflow. However, the system's success relies heavily on the 'compliance' of the AI agent. While adding instructions to AGENTS.md is a clever hook, the effectiveness of the knowledge base depends on the agent's discipline in running `gn help review`. If the agent glosses over the review phase, the knowledge base becomes sparse. Furthermore, as the JSONL files grow, the noise-to-signal ratio in search results may increase, potentially requiring a more sophisticated ranking mechanism than simple FTS. This tool is essential for teams moving toward 'agentic' workflows where AI agents are doing more than just autocomplete. If you are using Aider, Cursor, or Claude Code on a long-lived project, Gnosis provides a critical audit trail of intent that code alone cannot convey. It turns a series of ephemeral chat sessions into a permanent technical asset.