Coding agent memory layer released

Fullerenes is a persistent memory layer that scans repositories with Tree-sitter to extract functions, classes, imports, and call relationships, storing them in a local SQLite knowledge graph so coding agents don't repeatedly read the same files.

Agents connected via MCP query the graph instead of raw files. There are two core features.

• predict_impact returns direct and indirect dependencies and a risk score before a modification.
• get_function retrieves the signature, body, and callers all at once.

Benchmarks showed a significant reduction in token usage.

• SWE-bench Verified (1 case): reduced from 91,949 tokens to 32,945 tokens, a 64% reduction
• Internal repository (5 questions): reduced from an average of 2,452 tokens to 137 tokens, a 94.4% reduction
• External Python project (Gemini CLI): reduced from 27,292 tokens to 919 tokens, a 96.6% reduction

Because it's Tree-sitter-based, it captures structural relationships well but has weak semantics, so edges may be missed in code with heavy dynamic dispatch or metaprogramming.