Memorable

An agent finishes a task; Memorable stores how it was done and replays it when a similar task returns. Storage stays on your side and consent is fail-closed.

$npx memorable-cli@latest login

$memorable enable

$memorable recall "fix the tests"

$ npx memorable-cli@latest login

$ memorable enable

$ memorable recall "fix the tests"

Store your first procedure

Claude CodeCodexCursoryours

Wire up your harness

Claude Code, Codex, gbrain, QM, or yours. The harness field is any string.

Claudememorable

Connect Claude

One link, pasted into Claude. No terminal, nothing installed.

$ memorable recall "auth tests"

exactmiss

lexical0.86 hit

semantic0.79

How recall finds it

Exact, then lexical, then a vector. A reworded ask still lands.

POST /v1/extract

tool_calls[]

→

procedure

steps[]

One endpoint

POST the tool calls, get a procedure back. Same trace, same steps.

memorable init this machine

memorable init gbrain your gbrain db

memorable init qm QM postgres

Keep the store yours

This machine, your gbrain database, or QM Postgres.

agent turns-19%

What it saves

19% fewer agent turns, replicated.

Quickstart

Two minutes. Signing in links this machine to your workspace; the default backend is a standalone local store, and on a machine running gbrain you can point it at your own database with memorable init gbrain afterwards.

connect this machine

>npx memorable-cli@latest login# opens a browser, approves this machine

>memorable enable# explicit write consent, nothing is stored before this

no node or npm on this machine

>curl -fsSL https://memorable.sh/install.sh | sh

># then: memorable login

store a procedure, then find it

>memorable ingest trace.json

>memorable recall "rotate the TLS cert"

>memorable show <slug>

# guarded, injection-safe rendering

>memorable list

# everything stored, newest first

For agents

Telling a coding agent "use memorable" is enough. Every step is a plain CLI call it can run itself. The drop-in file lives on AGENTS.md.

one line

>memorable agents-md >> AGENTS.md

Measured

Against the same tasks run without memory, on two independent runs.

Recall vs no recall

agent turns   -19%, replicated
pass rate     every run passed
injected size ~293 tokens
vs a 15,593-token skill: 0% turns

Extraction API

One endpoint converts a tool-call trace into a procedure. Full reference on the API page.

POST /v1/extract

BASE=https://memorable-\
extraction-api.memorable.workers.dev

curl $BASE/v1/extract \
  -H "Authorization: Bearer \
$MEMORABLE_API_KEY" \
  -H "Content-Type: \
application/json" \
  -d '{
    "session_id": "run-183",
    "task_description": 
"rotate the TLS cert",
    "harness": "my-orchestrator",
    "tool_calls": [
      {"name": "shell",
       "input": {"command": 
"certbot renew"},
       "result": {"ok": true}}
    ]
  }'

Response · 200

{
  "draft": {
    "title": "Rotate the TLS cert",
    "schema_version": "1.0.0",
    "steps": [
      {
        "seq": 1,
        "action": "shell",
        "activity_class": "execute",
        "command": "certbot renew",
        "repeat_count": 1
      }
    ],
    "postconditions": [
      "final command exited 
successfully: certbot renew"
    ],
    "embedding": [],
    "embedding_model": ""
  },
  "request_id": "82886df0-..."
}