Memorable

+ x X 8 0 @# % $ S

# # @ @0 0 8 X

+ + x xX X 8 0

8 0 S S@ # % $

+ + x x 80 S S #

SCROLL TO CONTINUE

How It Works

LAYER 4Retrieval

The layer agents interact with.

When an agent starts a task, Memorable automatically finds the relevant workflow and guides the agent through it.

LAYER 4- Retrieval

LAYER 3Graph Assembly

Connect every workflow.

Shared steps and prefixes reveal new paths, allowing Memorable to compose workflows and learn implicit skills.

LAYER 3- Graph Assembly

LAYER 2Workflow Synthesis

Turn a successful run into a reusable workflow.

Memorable removes dead ends and keeps the sequence of steps that reached the goal.

LAYER 2- Workflow Synthesis

LAYER 1Traces

The raw record of an agent session.

A prompt, followed by the tool calls it produced: files read, commands run, and results returned, in order.

LAYER 1- Traces

Set Up In Three Steps.

npx memorable-cli

memorable 0.5.18

procedural memory for agents, stored on your own machine

backend gbrain · your own gbrain database

extraction configured · no model in the loop

commands

setup init setup login install-hooks agents-md

memory record ingest backfill recall show list chain

upkeep prune enable disable forget flush

inspect status doctor eval notices

start

RUNmemorable start

Sets all of this up for you.

{ PROCEDURE }

Bash

Reused most

$ memorable login

# opens a browser, approves this machine

$ memorable install-hooks

# recall runs on every new prompt

$ memorable enable

# explicit write consent, nothing stored before this

$ memorable recall "fix the failing auth test"

0.866 procedures/89f11bab-fix-failing-auth-test [lexical,semantic]

$ memorable show procedures/89f11bab-fix-failing-auth-test

THE FIX LANDED IN: tests/auth/session.test.ts

Verified last time by: npm test -- auth

$ memorable disable

# memory goes read-only. Recall still works, nothing new is recorded.

Install in one command

Point it at Claude Code, Codex, Cursor, or any other harness you run.

Set up once

Run the commands on the left to set up and sign in. Memorable then works in the background.

Create new memorables

Capture every workflow your agents run or recall as a reusable memorable.

Install in one command

Point it at Claude Code, Codex, Cursor, or any other harness you run.

Set up once

Run the commands on the left to set up and sign in. Memorable then works in the background.

Create new memorables

Capture every workflow your agents run or recall as a reusable memorable.

Reuse The Runs
That Worked.

Recall in ~60ms

Three matchers at once: exact, lexical, semantic.

RECALL TIME60msexact match

What is stored

The files touched, the command that proved it worked, and the order, with real exit codes. The transcript is never sent.

procedure_a91.json

files touchedgateway.py

verifypytest test.py

exit code0

order1 → 2 → 3

Works with your agent

Claude Code, Codex, OpenCode and gbrain have curated tool registries. Any other harness sends one JSON trace.

Chained plans

When a task is several things, recall returns an ordered plan of procedures and names any gap.

Savings per run

Tokens, money and time, next to the run they came from, including the runs where memory did not help.

Savings

vs. cold runs

Time Saved

R1R2R3R4R5R6R7R8R9R10R11R12

Scrubbed identifiers

Only the prompt and allow-listed tool arguments leave the machine.

local_trace.pyraw_session.json

t.session_id-"sess_8f2e1a9c"

t.user_email-"jordan@acme.com"

t.cwd-"/Users/jordan/billing-svc"

t.tool_calls-[ ... raw args ]

t.prompt-"process a customer refund"

STAYS ON THIS MACHINE

network_payload.pyreceived.json

payload = {

"prompt" : "process a customer refund",

"tool_args" : {

"command" : "pytest tests/refund",

"file_path" : "billing/refund.py"

}

# session_id, user_email, cwd,
# file contents, results: never sent

ONLY THIS LEAVES

Survives a tool changing shape

Procedures name a target by what it is about. In the cross-tool test every run hit the fastest path on a tool with no shared vocabulary.

memorable recall

$ memorable recall "process a customer refund"

# matched on shared tokens, not field names

0.86 procedures/9f21ac4d-process-a-customer-refund [lexical]

$ memorable show <slug>

# steps, files touched, verify command

One Memory For
Every Agent Type

For anyone running agents that do the same work more than once.

Browser use-

Known segments of a site replay from the procedure. A layout change is detected and the agent falls back to reasoning.

Computer use+

Each step stores only what changed on screen. The target is named by what it is about, so a moved button still matches.

Voice agents+

Routines taught in past sessions come back whole. Sensitivity gates at delivery, so the agent can know things it never says.

Coding agents+

Procedures come from Claude Code and Codex sessions, no code changed. The command that verified the fix travels with the procedure.

Other/custom/enterprise+

Any harness sends one JSON trace to POST /v1/extract. Store it on your own machine, your gbrain database, or your Postgres. Consent is fail-closed.

Multiagent Orchestration+

A procedure one agent recorded is recalled by every other agent on the same store.

01 / 06

Case Studies From Design Partners.

OpenHome

100% task completion, up from 70%

Every task finished with procedural memory. No wasted tool calls.

OpenHome

gstack

98% less context per prompt

The learned procedure is 293 tokens against 15,593 for the /investigate skill, on Claude Code with gbrain memory.

gstack

Quartermaster

40% fewer tool calls

5 to 3 tool calls on the same tasks. Pass rate 91% with memory, 80% without.

Quartermaster

OpenHome

100% task completion, up from 70%

Every task finished with procedural memory. No wasted tool calls.

gstack

98% less context per prompt

The learned procedure is 293 tokens against 15,593 for the /investigate skill, on Claude Code with gbrain memory.

Quartermaster

40% fewer tool calls

5 to 3 tool calls on the same tasks. Pass rate 91% with memory, 80% without.

Benchmarks

GBrain, Claude Code

TURNS PER TASK, 3-BUG CODING FIXTURE

-19%

Without memory16

With Memorable13

454 runs, every one passed.

Quartermaster, Codex

TOOL CALLS PER TASK, SAME FIXTURE

-40%

Without memory5

With Memorable3

Three replications, same result each time.

Frequently Asked Questions

Does Memorable work with my agent?

Claude Code, Codex, Cursor, Claude Cowork, gbrain, and QM. Any other harness sends one JSON trace to POST /v1/extract. A read-only MCP server exposes five tools.

How does procedural graph memory differ from vector search?

Vector search returns chunks of text. Memorable stores procedures: steps with a precondition, a postcondition, and the outcome that decided whether the run passed. The step parse is deterministic: the same trace gives the same steps.

Is data stored locally, self-hosted, or cloud-managed?

Procedures live in your own store: a local file, your gbrain database, or your Postgres. Embeddings are computed server-side. Only the prompt and allow-listed tool arguments leave the machine.

What is the latency overhead on agent inference calls?

Recall adds about 60 milliseconds to a prompt. Three matchers run in order: exact, then lexical, then semantic on a miss.

Can we inspect and visualize the agent memory graph?

The dashboard shows every stored procedure: steps, preconditions, postconditions, and the time, tokens, and money spent on each run.

Agents That Get Better With Every Run.

Procedural, Graph Based
Memory For Agents.

How It Works

Set Up In Three Steps.

Install in one command

Set up once

Create new memorables

Install in one command

Set up once

Create new memorables

Reuse The Runs
That Worked.

Recall in ~60ms

What is stored

Works with your agent

Chained plans

Savings per run

Scrubbed identifiers

Survives a tool changing shape

One Memory For
Every Agent Type

Case Studies From Design Partners.

100% task completion, up from 70%

98% less context per prompt

40% fewer tool calls

100% task completion, up from 70%

98% less context per prompt

40% fewer tool calls

Benchmarks

GBrain, Claude Code

Quartermaster, Codex

Frequently Asked Questions

Does Memorable work with my agent?

How does procedural graph memory differ from vector search?

Is data stored locally, self-hosted, or cloud-managed?

What is the latency overhead on agent inference calls?

Can we inspect and visualize the agent memory graph?

Let's Build Something
Memorable

Agents That Get Better With Every Run.

How It Works

Set Up In Three Steps.

Install in one command

Set up once

Create new memorables

Install in one command

Set up once

Create new memorables

Reuse The Runs That Worked.

Recall in ~60ms

What is stored

Works with your agent

Chained plans

Savings per run

Scrubbed identifiers

Survives a tool changing shape

One Memory ForEvery Agent Type

Case Studies From Design Partners.

100% task completion, up from 70%

98% less context per prompt

40% fewer tool calls

100% task completion, up from 70%

98% less context per prompt

40% fewer tool calls

Benchmarks

GBrain, Claude Code

Quartermaster, Codex

Frequently Asked Questions

Does Memorable work with my agent?

How does procedural graph memory differ from vector search?

Is data stored locally, self-hosted, or cloud-managed?

What is the latency overhead on agent inference calls?

Can we inspect and visualize the agent memory graph?

Let's Build SomethingMemorable

Reuse The Runs
That Worked.

One Memory For
Every Agent Type

Let's Build Something
Memorable