Design Notes

Design spikes and data-model write-ups for repo-memory’s bigger subsystems, ported from the repo’s docs/planning/ directory. These capture the reasoning at design time — schemas, algorithms, tradeoffs — and are kept as design history; where the shipped implementation diverged, each page notes it. Source: github.com/blamechris/repo-memory.

Pages

• AST Summarizer — the tree-sitter (WASM) spike behind "summarizer": "ast": dependency choice, accuracy and speed measurements, generation-tag cache invalidation, and the rollout from TS/JS to Python/Go/Rust and Kotlin/Java.
• Task Memory Data Model — task lifecycle, the two-table schema, the computed frontier, and GC rules behind create_task / get_task_context / mark_explored.
• Dependency Graph — SQLite adjacency-list storage, in-memory query patterns, incremental updates, and multi-language extensibility behind get_dependency_graph.
• Relevance Ranking — the weighted-signal scoring model behind get_related_files, including the 0.13.0 rework (relationship and centrality signals; the shipped weights).
• Diff-Aware Summary Updates — classifying changes as structural vs non-structural via git diff to skip unnecessary re-summarization; designed and deferred until scale warrants it.
• Agent Search Efficiency Audit — the six-lens June 2026 audit of the retrieval path: the cache-poisoning bug, the write-only graph, the token-shaping pass, and the explicit non-goals (embeddings, query caching, new tools; FTS5 deferred). Everything in its action plan shipped in 0.13.0.

See Architecture and Caching for how the shipped subsystems fit together, and the tools reference for the tools they power.