HN 표시: PgVector 및 BM25를 사용하여 LongMemEval에서 81% – Claude 코드용 메모리

Verbatim semantic memory for Claude Code sessions. Every prompt, every response, every project — searchable by meaning, across every session you've ever run. Self-hosted. Free. ~10 minutes to install. "I ran three sessions yesterday on the same project. Which one fixed the bug?" Echoes answers that question in 200ms, in plain English, from your verbatim history. Benchmarked honestly. 81.0% on LongMemEval (ICLR 2025) with Sonnet 4.6 — pgvector cosine + BM25 RRF hybrid + temporal re-ranking. 100% on single-session-user retrieval (70/70). Full per-category breakdown, raw outputs, and reproduction steps in benchmarks/. No hardcoded answer patterns. No invented terminology. No cherry-picking. Claude Code is great at one session. It has no idea what happened in the last one. Echoes captures every user prompt and assistant response as you work, stores them in Postgres with pgvector embeddings from a local Ollama model (nomic-embed-text , 768-dim), and exposes a /recall skill inside Claude Code so you can ask: /recall when did we fix the partition key bug /recall cors issue with azure functions --project azureprep /recall vercel deployment failure --days 7 and get the actual verbatim conversation back — with project, role, date, and a session ID so you can pull the surrounding context. No summarisation. No LLM deciding what's "important." No vendor lock-in. No cloud. Your data stays on your machine (or your VPS). The embedding model is a 137M-parameter local model. Retrieval is a single SQL query against an HNSW index. The entire thing is ~600 lines of Python + 200 lines of JavaScript. Honest comparison: | claude-echoes | MemPalace | mem0 / letta | | |---|---|---|---| | Claude Code hook integration | Native, single-file JS hook | No | No | | Storage | Postgres + pgvector | ChromaDB | Varies | | Embedding model | Local Ollama (free, no API) | Local | Usually OpenAI | | Cross-session semantic recall | Yes | Yes | Yes | | Install time | ~10 min (one docker-compose) | ~30 min | Varies | | Designed for Claude Code specifically | Yes | No | No | | Summarises / extracts "important" bits | No — verbatim only | No | Yes | | LongMemEval_S score | 81.0% (Sonnet) / 64.4% (Haiku) — reproducible | 96.6% claimed (disputed) | Not reported | | Lines of code | ~800 + ~600 benchmark | ~3000+ | Large | If you already run a Postgres and don't use Claude Code, MemPalace is probably better for you. If you live in Claude Code all day and keep losing context between sessions, this is built for you. Requires: Docker, a Claude Code install, and either Linux, macOS, or Windows (WSL2 or Git Bash). git clone https://github.com/m4cd4r4/claude-echoes.git cd claude-echoes ./scripts/install.sh The installer will: - Spin up Postgres 15 + pgvector and Ollama via docker-compose - Pull the nomic-embed-text model (~275MB) - Apply the schema migration - Install the Claude Code hook into ~/.claude/hooks/claude-echoes/ - Install the /recall skill into~/.claude/skills/claude-echoes/ - Print a test command to verify Total time: ~10 minutes (most of which is pulling the ollama model). After install, just use Claude Code normally. Every prompt and response is captured and embedded in the background with zero added latency. Within a few sessions, /recall starts returning useful results. Inside any Claude Code session: /recall Filters: /recall --project # limit to one project /recall --days # only messages from last N days /recall --role user|assistant # filter by sender /recall --limit # how many hits (default 10) Or call the HTTP API directly: curl "http://localhost:8088/search?q=how+did+we+fix+the+auth+bug&limit=5" ┌─────────────────┐ │ Claude Code │ │ (your IDE) │ └────────┬────────┘ │ UserPromptSubmit / Stop hook ▼ ┌─────────────────┐ │ chat-logger.js │ (hooks/chat-logger.js, ~150 LOC) └────────┬────────┘ │ POST /message ▼ ┌─────────────────┐ │ echoes-server │ (server/app.py, ~200 LOC) │ FastAPI │◄──────┐ └────┬───────┬────┘ │ │ │ │ embed on insert│ │search │GET /search?q=... ▼ ▼ │ ┌──────────┐ ┌──────────┐ │ │ Ollama │ │ Postgres │ │ │ nomic- │ │ + HNSW │ │ │ embed │ │ pgvector │ │ └──────────┘ └──────────┘ │ │ ┌─────────────────┐ │ │ /recall skill │───────┘ │ (SKILL.md) │ └─────────────────┘ Flow on every message: - Claude Code fires UserPromptSubmit orStop hook event chat-logger.js POSTs the message tohttp://localhost:8088/message - Server calls Ollama to embed the content (~150ms, in-process) - Server inserts (session_id, project, role, content, embedding) into Postgres - Total latency: ~200ms, fully async, never blocks your prompt Flow on /recall : - You type /recall in any session - Skill calls GET /search?q=&...filters - Server embeds the query, runs ORDER BY embedding $1::vector LIMIT N against the HNSW index - Returns ranked hits with full verbatim content + metadata - Claude renders them with similarity scores and project/date context - It does not summarise. Your history is verbatim. If you want a summary, ask Claude to summarise the recall resul