quietmem
Local, precision-first memory for AI coding agents: recalls past conversations only when it is confident they matter. Claude Code hooks + MCP server, pgvector + Ollama, no API keys.
Safety scan
A static text scan, not a guarantee. Select the badge to see what was scanned and every finding.
Badge for your README
Markdown
[](/listings/ronto1138--quietmem)HTML
<a href="/listings/ronto1138--quietmem"><img src="/api/listings/ronto1138--quietmem/badge.svg" alt="findsafeskills safety rating" /></a>Install
Install instructions from the README
As written by the repo's author. findsafeskills has not verified these steps.
Requirements: Node 20+, Docker (or any Postgres 15+ with pgvector), and [Ollama](https://ollama.com).
git clone https://github.com/ronto1138/quietmem && cd quietmem && npm install
ollama pull nomic-embed-text:v1.5
export QUIETMEM_DB_PASSWORD="$(openssl rand -hex 16)"
docker compose up -d
mkdir -p ~/.quietmem && cp quietmem.config.example.json ~/.quietmem/config.json # edit it: entities, sources
export QUIETMEM_DATABASE_URL="postgresql://quietmem:${QUIETMEM_DB_PASSWORD}@localhost:5433/quietmem"
node bin/quietmem.js init # tables and indexes
node bin/quietmem.js index --sweep # index your existing Claude Code transcripts
node bin/quietmem.js tf # term frequencies used by the precision filters
node bin/quietmem.js doctor # checks everything and times each phaseKeep `QUIETMEM_DATABASE_URL` in your shell profile (not in the config file), so the password never lands in a file you might share.
Claude Code
node bin/quietmem.js install claude-code # prints the hooks to add to ~/.claude/settings.json node bin/quietmem.js install claude-code --write # or merges them for you (keeps a backup)
That adds two hooks: `UserPromptSubmit` runs recall, and `Stop` indexes what was just said, in the background.
Any MCP client
{ "mcpServers": { "quietmem": { "command": "node", "args": ["/path/to/quietmem/bin/quietmem.js", "mcp"],
"env": { "QUIETMEM_DATABASE_URL": "postgresql://quietmem:...@localhost:5433/quietmem" } } } }Tools: `search_memory` (plain search the model can call when the user refers to earlier work) and `recall_memory` (the same gated recall the hooks use). Clients without hooks cannot inject automatically, so indexing their conversations needs an adapter or an export to the [JSONL format](docs/adapters.md).
What was scanned
The files the safety scan read, as written by the repository's author and shown as plain text. Lines with a finding are marked.
README.md · README · 102 lines
- # quietmem
- **Local, precision-first memory for AI coding agents.** quietmem indexes your past conversations with your
- agents and, when a new message clearly relates to something you discussed before, quietly hands the
- relevant fragments to the model. When it is not confident, it injects nothing.
- - **Everything local.** Postgres + pgvector and a local embedding model (Ollama by default). No API keys
- needed, no cloud, no telemetry.
- - **No resident process.** Hooks start, do their work and exit. Recall has a hard time budget (500 ms by
- default): if it cannot answer in time, the message goes through untouched.
- - **Precision over recall.** Similarity alone is a bad signal: "yes, go ahead" scores 0.74 against half
- your history. quietmem requires the message to name something known (a project, a client, a tool) or to
- ask to remember, and every fragment must share distinctive terms with it. See [docs/design.md](docs/design.md).
- - **Secrets stay masked.** Transcripts contain pasted keys. Everything quietmem prints goes through a redactor.
- - **Memory you approve.** An optional nightly job reads the day's conversations and *proposes* durable
- memories (decisions, preferences, references) as Markdown files. Nothing is written until you accept it.
- - **Any harness, any model.** Hooks for Claude Code; an MCP server for Cursor, Codex CLI, Windsurf, Zed,
- Claude Desktop and any other MCP client; a plain JSONL format for everything else. Embeddings from
- Ollama or any OpenAI-compatible endpoint; proposals from whatever command you configure (`claude -p`,
- `codex exec`, `ollama run`, your own script).
- ## Why
- quietmem started as a replacement for a third-party memory plugin that injected too much, ran a resident
- server, and quietly spent a lot of API tokens. The goal was the opposite on every axis: local, silent when
- unsure, measurable, and free to run. It has been in daily use across a dozen projects since September 2026,
- on a corpus of 50k+ chunks in Spanish and English.
- ## Quick start
- Requirements: Node 20+, Docker (or any Postgres 15+ with pgvector), and [Ollama](https://ollama.com).
- ```bash
- git clone https://github.com/ronto1138/quietmem && cd quietmem && npm install
- ollama pull nomic-embed-text:v1.5
- export QUIETMEM_DB_PASSWORD="$(openssl rand -hex 16)"
- docker compose up -d
- mkdir -p ~/.quietmem && cp quietmem.config.example.json ~/.quietmem/config.json # edit it: entities, sources
- export QUIETMEM_DATABASE_URL="postgresql://quietmem:${QUIETMEM_DB_PASSWORD}@localhost:5433/quietmem"
- node bin/quietmem.js init # tables and indexes
- node bin/quietmem.js index --sweep # index your existing Claude Code transcripts
- node bin/quietmem.js tf # term frequencies used by the precision filters
- node bin/quietmem.js doctor # checks everything and times each phase
- ```
- Keep `QUIETMEM_DATABASE_URL` in your shell profile (not in the config file), so the password never lands in a
- file you might share.
- ### Claude Code
- ```bash
- node bin/quietmem.js install claude-code # prints the hooks to add to ~/.claude/settings.json
- node bin/quietmem.js install claude-code --write # or merges them for you (keeps a backup)
- ```
- That adds two hooks: `UserPromptSubmit` runs recall, and `Stop` indexes what was just said, in the background.
- ### Any MCP client
- ```json
- { "mcpServers": { "quietmem": { "command": "node", "args": ["/path/to/quietmem/bin/quietmem.js", "mcp"],
- "env": { "QUIETMEM_DATABASE_URL": "postgresql://quietmem:...@localhost:5433/quietmem" } } } }
- ```
- Tools: `search_memory` (plain search the model can call when the user refers to earlier work) and
- `recall_memory` (the same gated recall the hooks use). Clients without hooks cannot inject automatically,
- so indexing their conversations needs an adapter or an export to the [JSONL format](docs/adapters.md).
- ## Commands
- | Command | What it does |
- |---|---|
- | `quietmem init` | Create tables and indexes (dimensions and index type from the config) |
- | `quietmem doctor` | Check config, embedder and database; time a cold and a warm vector query |
- | `quietmem index --sweep` | Index everything new in the configured sources |
- | `quietmem index <file> [--adapter jsonl]` | Index one transcript |
- | `quietmem tf` | Rebuild term frequencies (after big corpus changes) |
- | `quietmem recall "text" [--debug]` | See what the gates would do with a message |
- | `quietmem search "text" [--project p]` | Plain semantic search, no gates |
- | `quietmem mcp` | MCP server over stdio |
- | `quietmem distill [--date d] [--project p]` | Propose memories from a day of conversation |
- | `quietmem proposals expire` | Move stale pending proposals to `expired/` |
- ## Documentation
- - [Configuration](docs/configuration.md): every option, entities, project labels, embedders
- - [Design](docs/design.md): the gates, why they exist, how to tune them on your corpus
- - [Adapters](docs/adapters.md): add support for another harness (the best first contribution)
- - [Performance](docs/performance.md): the recall budget and keeping the index warm
- ## Status
- v0.1. Used daily by its author; the generalized packaging is new. Expect rough edges, and please open
- issues. Contributions are very welcome, especially adapters for other harnesses and calibration results
- on corpora in other languages: see [CONTRIBUTING.md](CONTRIBUTING.md).
- ## License
- [MIT](LICENSE)
Ratings and comments
Sign in to rate this skill.
Sign inFree to use. If it’s helped, pass it on.