Steno – AI 에이전트용 RAG를 사용한 압축 메모리

Compressed memory notation with RAG retrieval for AI agents. Steno solves the AI memory problem: agents accumulate knowledge across sessions, but loading everything into context every time is expensive, noisy, and causes drift. Steno compresses memories into a dense notation format and retrieves only what's relevant using semantic search. AI coding agents (Claude Code, Cursor, Copilot) build up memory files over time — user preferences, project context, past decisions, feedback. The default approach is brute-force: load all memory into every session. This wastes tokens, pollutes context with irrelevant information, and causes the agent to act on stale facts. Two-tier notation: - Steno — human-auditable compressed format. Drop articles, abbreviate common terms, use key-value pairs. Readable by humans, efficient for AI. - Steno-M — AI-only format. Fixed schemas, positional fields, no labels. Maximum density for machine-to-machine communication. RAG retrieval: - Parse memory files into structured records - Embed with a lightweight model (all-MiniLM-L6-v2, 80MB, runs on CPU) - Store in ChromaDB (local, no server needed) - Query semantically — only relevant memories enter the context window # Clone git clone https://github.com/YOUR_USERNAME/steno.git cd steno # Install python3 -m venv .venv source .venv/bin/activate pip install -r requirements.txt # Index your memory files python steno.py index ./examples # Search python steno.py query "database testing rules" 5 # Check index stats python steno.py stats Memory files are markdown with YAML frontmatter: --- name: Integration Tests Must Hit Real DB description: No mocking database in integration tests type: feedback format: steno --- rule: integration tests connect to real PostgreSQL, never mock DB layer **Why:** Q1 migration failure — mocked tests passed but prod migration broke. **How to apply:** tests/integration/ → always uses test_db fixture (real PostgreSQL) tests/unit/ → mocks are fine (testing logic, not persistence) | Type | What | When to Save | |---|---|---| user | Role, goals, preferences | When you learn about the user | feedback | Corrections, confirmed approaches | When the user corrects or validates | project | Ongoing work, decisions, deadlines | When you learn project context | reference | Pointers to external resources | When you discover useful external info | - Drop articles (a, the, an) - Abbreviate common terms: verified →vfd ,authentication →auth ,configuration →cfg - Compress dates: 2026-04-11 →04-11 - Key-value for metadata: key: value - Indentation for hierarchy - Full words for anything ambiguous For machine-to-machine communication, Steno-M uses fixed schemas with positional fields: #scope myproject #schemas @F @T @F BUG-123|open|high|auth-bypass-on-reset|auth-service|evidence/bug123 @T auth-service|active|go-grpc|primary auth, port 50052 @T billing-api|migrating|python-rest|gRPC migration 60% Record types: @V (vulnerability), @F (finding), @T (target), @C (credential), @A (message), @L (lead) steno index [--rebuild] MEMORY_DIR Index memory files (incremental by default) steno query "text" [N] Semantic search, top N results steno stats Show index statistics steno parse FILE_OR_DIR Parse and preview records # Filter by memory type steno query "deployment process" --memory=project # Filter by record type (Steno-M) steno query "auth service" --type=@T # Set minimum similarity score steno query "testing" --min=0.6 | Variable | Default | Description | |---|---|---| STENO_STORE | ./chroma_store | ChromaDB storage path | STENO_MODEL | all-MiniLM-L6-v2 | Sentence-transformers model | Memory files (Steno/Steno-M/prose) ↓ parse (steno_parser.py) Structured records with metadata ↓ embed (memory_index.py) ChromaDB vector store (local) ↓ query (memory_retrieval.py) Top-K relevant records ↓ inject AI agent context window Incremental indexing: Steno tracks file modification times. Only changed files are re-embedded on re-index. Unchanged files are skipped in ~0.1s. Junk filtering: Empty lines, separators (--- ), and content-free fragments are automatically filtered out during parsing. from memory_index import build_index from memory_retrieval import query # Index build_index(Path('./memories')) # Query results = query("what are the deployment rules", top_k=5) for r in results: print(f"[{r.score:.2f}] {r.source_file}: {r.text[:100]}") Place your memory files in ~/.claude/projects/YOUR_PROJECT/memory/ and run: steno index ~/.claude/projects/YOUR_PROJECT/memory/ steno query "relevant context for current task" The query_formatted() function returns a pre-formatted string ready for context injection: from memory_retrieval import query_formatted context = query_formatted("auth service architecture", top_k=5) # Returns formatted block ready to inject into agent prompt Tested on 85 memory files (466 KB total): | Operation | Time | |---|---| | Full rebuild | ~20s | | Incremental (no changes) | 0.08s | | Single file re-index | ~3s | | Query | ~2s (includes model load) | Model: all-MiniLM-L6-v2 (384-dim, 80MB, CPU-only). No GPU required. MIT