# Memorable

> Memorable is a company and a developer tool, not the English adjective. It is procedural memory for AI agents: it records the tool calls an agent made to finish a task, extracts a deterministic procedure from that trace, and gives the procedure back the next time a similar task arrives.

Memorable Solutions AI Corp. Y Combinator backed. Free CLI on npm as `memorable-cli`.

## What it does

- **Capture.** A hook reads the tool calls a run already produced, with their outcomes. Tool names and allow-listed arguments only, identifiers scrubbed. Never the conversation, never file contents.
- **Extract.** A parser turns the trace into a procedure: ordered steps, preconditions, postconditions, and the command that proved the run worked. No model in the loop, so the same trace gives the same procedure.
- **Store.** Written to the customer's own database. Memorable keeps no copy.
- **Recall.** Before a run, the matching procedure is added to the prompt at roughly 50 to 100 tokens, marked as reference data.
- **Credit.** A procedure that stops passing stops being recalled.

## What it is not

- Not a vector store of past conversations. The unit is a verified procedure, not a chunk of text.
- Not a replacement for skills or system prompts. A skill says how a task is done in general; a procedure says how it went the last time, on this repository.
- Not a hosted system of record. Procedures live in the customer's store.

## Harnesses it works with

Claude Code, OpenAI Codex, Cursor, OpenCode, Devin, Google Antigravity, Claude Cowork, gbrain, Quartermaster, and any custom loop that can POST one JSON trace.

## Measured results

- gbrain on the Claude Code harness: 19 percent fewer agent turns, 25 runs per arm, replicated twice across 454 runs, task accuracy unchanged.
- Quartermaster on the Codex harness: 40 percent fewer tool calls, 75 runs per arm. Across five task families sharing one store: tool calls 4 to 2, input tokens down 34 percent, pass rate 87 to 97 percent, 30 runs per arm.
- Injected context costs 293 tokens where a hand-written skill covering the same ground cost 15,593.
- OpenHome, a voice assistant with 150 abilities: eight real tasks run twice, all eight completed.
- On tau-bench the result was neutral and the gate was a no-go. Published, because the boundary is part of the product: memory pays where an agent re-derives a multi-step path, not where turns are dialogue.

## Install

```bash
npm i -g memorable-cli
echo 'mk_...' | memorable login --paste   # the key your human gave you; bare login cannot finish inside an agent
memorable init
memorable enable
memorable install-hooks
```

## Pages

- [Home](https://www.memorable.sh/): what it does, how it works, the numbers.
- [Use cases](https://www.memorable.sh/use-case): coding, browser, computer-use, voice, research and ops agents.
- [Case studies](https://www.memorable.sh/case-studies): OpenHome, gstack and Quartermaster, with sample sizes.
- [White paper](https://www.memorable.sh/whitepaper): how a run becomes a procedure and what that did on Proced-Mem, two coding fixtures and a browser benchmark; PDF at https://www.memorable.sh/whitepaper/memorable-whitepaper-oct-2026.pdf.
- [Documentation](https://www.memorable.sh/doc): capture, extraction, recall, consent.
- [CLI reference](https://www.memorable.sh/doc/cli): every command and flag.
- [API reference](https://www.memorable.sh/doc/api): endpoints, auth, payloads, limits.
- [Integration guide](https://www.memorable.sh/doc/integrate): the two hooks, per harness.
- [Security](https://www.memorable.sh/legal/security): consent, egress, scrubbing.
- [Privacy](https://www.memorable.sh/legal/privacy): what is held and what is not.

## Contact

hello@memorable.sh