DMCA is a symbolic semantic memory system designed to complement and extend traditional vector databases. It serves as the foundation for FixCom's intelligent, AI-assisted release note generator
DMCA is to power the backend of FixCom, a tool that automates Git commit summarization, changelog generation, and release note authoring using small quantized models.
- Parses commits into hunks
- Summarizes each hunk independently using tiny LLMs (less context rot)
- Aggregates hunks into commit-level summaries
- Tags hunks using Ctags, and stores keywords in a vector DB
- Segments release notes by gathering hunks related to a given keyword
- However, this system is limited: the vector DB lacks semantic understanding, relationship modeling, and context merging.
DMCA addresses the limitations of static keyword search by constructing a symbolic Markov graph, where:
- Nodes represent source-level entities (e.g., functions, classes, variables) and semantic concepts (e.g., "thread-safe", "render pipeline").
- Edges represent probabilistic transitions between symbols based on co-occurrence, file proximity, call relations, or summarizer keywords.
- Commits are modeled as subgraphs, giving them semantic “footprints.”
- Features (like threading changes, renderer refactors, etc.) are inferred as connected regions of the graph.
- Discover semantically related commits automatically
- Merge related concepts and track refactors over time
- Build structured, modular release notes per feature
- Provide full explainability without large LLMs
Git Repo
│
▼
[ FixCom Hunks ]
│
+-----┴------+-----------------------------+
| | |
Small Model ➜ Summary Extract CTags
| |
Commit Summary ◄───────────── Tag/Concept Mapping
| |
+-------------→ [ DMCA Graph ] ←----------+
│
▼
Graph Regions + Commit Subgraphs
│
▼
[ DMCA Output ]
│
▼
FixCom Generates Release Notes