repomemory 0.2.3

AI-powered code retrieval — index any GitHub repo by URL, search with natural language, get token-budget-aware context packs for LLMs

RepoMemory

AI-powered code retrieval engine — index any GitHub repo, search with natural language

Point it at any GitHub URL. Get token-budget-aware context packs — ready to paste into any LLM. Free to run, free to deploy.

---

What is RepoMemory?

When you ask an LLM to fix a bug or trace a feature, it needs the right source files. Pasting the whole codebase wastes the context window. Guessing which files to include misses critical pieces.

RepoMemory solves this with a hybrid retrieval pipeline that runs on any public or private GitHub repo:

repomemory index https://github.com/pallets/flask
repomemory search "Where is request routing handled?"

Query  →  Task Classification (trace_flow)
       →  BM25 Lexical + FAISS Semantic + Fuzzy Path + Symbol search  (parallel)
       →  Reciprocal Rank Fusion  →  Top-20 ranked files
       →  Dependency-graph expansion  →  Re-rank with adaptive weights
       →  Token-budget packer  →  Context pack ready to paste into any LLM
       →  (optional) Groq AI summary

---

Install

# Core — CLI + library, uses HuggingFace API for embeddings (no GPU needed)
pip install repomemory

# With local embeddings (~80 MB model download, fully offline)
pip install "repomemory[local]"

# With FastAPI web server
pip install "repomemory[server]"

# With Groq LLM explanations (free API key)
pip install "repomemory[llm]"

# Everything
pip install "repomemory[all]"

---

Quick Start

CLI

# Index a repo
repomemory index https://github.com/pallets/flask

# Search it
repomemory search "How does request routing work?"

# With AI explanations (free Groq key)
export REPOMEMORY_GROQ_API_KEY=gsk_...
repomemory search "Where is token rotation handled?"

# Adjust result count and token budget
repomemory search "auth middleware" --top-k 10 --budget 4000

# Force a task mode
repomemory search "test coverage for auth" --mode test_lookup

# Private repos
repomemory index https://github.com/myorg/private-repo --token ghp_...

# List indexed repos
repomemory list

# Start the web UI + API server
repomemory serve

Python library

from repomemory import RepoMemory

rm = RepoMemory()

# Index
repo = rm.index("https://github.com/pallets/flask")

# Search
result = rm.search("How does request dispatching work?")
for file in result.context_pack.files:
    print(f"{file.path}  score={file.relevance_score:.2f}")
    print(file.snippets[0].content[:300])

REST API

# Start the server
pip install "repomemory[server]"
repomemory serve

# Index a repo
curl -X POST http://localhost:8000/api/repos \
  -H "Content-Type: application/json" \
  -d '{"url": "https://github.com/pallets/flask"}'

# Search
curl -X POST http://localhost:8000/api/search \
  -H "Content-Type: application/json" \
  -d '{"repo_id": 1, "query": "How does routing work?", "token_budget": 8000}'

---

Features

Feature	Details
Hybrid search	BM25 lexical + FAISS semantic (384-dim) + fuzzy path + symbol name, fused with Reciprocal Rank Fusion
Dependency-graph retrieval	Builds file-level import edges at index time; expands results through related files via BFS
Adaptive weight learning	Online SGD learning from user feedback (accept / dismiss / thumbs); falls back to static mode weights
Symbol-aware indexing	tree-sitter extracts functions, classes, and methods from Python, JavaScript, and TypeScript
5 Task Modes	bug_fix, trace_flow, test_lookup, config_lookup, general — auto-detected from query or set manually
Token-budget packer	Greedy packer respects any token limit (default 8 000 tokens, configurable up to 100 000+)
Behavioral memory	Frecency scoring from opened / accepted / thumbs-up actions; boosts relevant files in future queries
RAG evaluation	End-to-end pipeline scoring retrieval impact on LLM answer quality (relevance, completeness, faithfulness)
Flexible embeddings	Local sentence-transformers (offline) or free HuggingFace Inference API
AI explanations	Optional Groq LLM (free tier) explains why each result matters
Incremental indexing	SHA-256 per file; only changed files are re-embedded on re-index
Web UI + REST API	React 19 frontend + FastAPI backend; deploy on Render + Vercel (both free)
Export as Markdown	Copy context pack as a formatted Markdown block to paste directly into an LLM prompt

---

Task Modes

RepoMemory classifies each query and adjusts retrieval weights automatically:

Mode	Auto-detected from	What it boosts
bug_fix	error, exception, crash, fix, traceback	Lexical signal, error-adjacent files
trace_flow	trace, flow, route, handler, how does...work	Symbol matching, call-chain ordering
test_lookup	test, spec, mock, fixture, coverage	Path matching for tests/ / spec/ dirs
config_lookup	config, env, setting, yaml, toml	Path matching for config-like files
general	(fallback)	Balanced across all signals

---

Configuration

All settings use the REPOMEMORY_ env prefix (powered by pydantic-settings):

export REPOMEMORY_HF_API_KEY=hf_...          # HuggingFace free API key
export REPOMEMORY_GROQ_API_KEY=gsk_...       # Groq free API key (for AI summaries)
export REPOMEMORY_EMBEDDING_PROVIDER=local   # 'local' or 'huggingface'
export REPOMEMORY_DATA_DIR=/data/repomemory  # where SQLite + FAISS live
export REPOMEMORY_TOKEN_BUDGET=16000         # default context pack size

---

Links

• GitHub — github.com/aayushakumar/RepoMemory (full docs, architecture diagram, contributing guide)
• Issues — github.com/aayushakumar/RepoMemory/issues
• Changelog — github.com/aayushakumar/RepoMemory/releases
• API Docs (when running locally) — http://localhost:8000/docs

---

License

MIT © Aayush Kumar