Memorable is procedural memory infrastructure for AI agents, aimed at teams running coding assistants, browser agents, voice agents, and custom harnesses that repeat similar work. The core idea is to capture a successful agent run as a graph of the tool calls that actually reached the goal, then replay that path the next time a similar task appears, instead of letting the agent rederive the plan from scratch inside a fresh context window. The company installs into Claude Code, Codex, Cursor, Claude Cowork, gbrain, and QM, and any other harness can send a single JSON trace to a POST /v1/extract endpoint, with a read-only MCP server exposing tools for recall.
Under the hood the product is layered from raw traces up through workflow synthesis and graph assembly to retrieval, where shared steps and prefixes across workflows reveal new paths, allowing Memorable to compose workflows and learn implicit skills . Storage stays local by design. Recall runs against three matchers at once (exact, lexical, and semantic), and procedures name their targets by intent rather than field names, which the team says lets a moved button or a renamed tool still match. Early design partners include OpenHome, gstack, and Quartermaster, with published case studies pointing at reductions in tool calls and context per prompt on repeat tasks.
The launch matters now because agent operators are hitting the ceiling of stuffing more skills, more instructions, and longer context into every run, and procedural memory is emerging as the alternative most production teams are exploring. Memorable is a Y Combinator S27 company built by Advaiyt Sane and Nikhil K (@nikhilk8754), and this launch opens up the CLI, hooks, and dashboard for teams that want their agents to keep what they learn across sessions rather than starting cold every time.
Comments (9)
Sign in to join the discussion.
N
Nneka O.25d ago
YC S27? Time traveling founders now get a batch label before the batch exists. The pre-crime of accelerators.
R
Raghav Menon25d ago
Graph based memory is cool but I need to know the eviction policy before I care. What happens when two successful runs contradict each other, last write wins or some voting thing?
T
toby25d ago
Agents that remember successful runs sounds great until they start cargo culting a fluke from three weeks ago into every workflow.
L
Leo da Cruz25d ago
The tweet cuts off mid sentence on 'reusable https://t.co/...' and honestly that might be the strongest hook I've seen this week. I clicked.
M
Marta Kowalczyk25d ago
Every agent memory startup pitches the same 'born, work, die' line. At this point it's less a thesis and more a genre convention.
H
Hans Olvig25d ago
Everyone's shipping memory layers and nobody's shipping a benchmark that isn't authored by the same team. Show me an eval you didn't write.
D
Deep Das25d ago
Been building something adjacent for 3 weeks, was going to launch next month. Guess I'm pivoting to procedural memory for procedural memory.
K
kestrel25d ago
A graph that learns from its own success is just a superstition with better tooling.
P
Priya S.25d ago
Landing page loads fast, tagline is legible, launch tweet has a clean cold open. Rare trifecta for infra launches this quarter.
Work With Us
Book an intro call to explore our suite of services. We might have a waitlist.
Memorable
Procedural, Graph Based Memory for agents
Procedural, Graph Based Memory for agents
Tags
TLVC Rating
Community Rating
No ratings yet
to rate this launch.
Memorable is procedural memory infrastructure for AI agents, aimed at teams running coding assistants, browser agents, voice agents, and custom harnesses that repeat similar work. The core idea is to capture a successful agent run as a graph of the tool calls that actually reached the goal, then replay that path the next time a similar task appears, instead of letting the agent rederive the plan from scratch inside a fresh context window. The company installs into Claude Code, Codex, Cursor, Claude Cowork, gbrain, and QM, and any other harness can send a single JSON trace to a POST /v1/extract endpoint, with a read-only MCP server exposing tools for recall. Under the hood the product is layered from raw traces up through workflow synthesis and graph assembly to retrieval, where shared steps and prefixes across workflows reveal new paths, allowing Memorable to compose workflows and learn implicit skills . Storage stays local by design. Recall runs against three matchers at once (exact, lexical, and semantic), and procedures name their targets by intent rather than field names, which the team says lets a moved button or a renamed tool still match. Early design partners include OpenHome, gstack, and Quartermaster, with published case studies pointing at reductions in tool calls and context per prompt on repeat tasks. The launch matters now because agent operators are hitting the ceiling of stuffing more skills, more instructions, and longer context into every run, and procedural memory is emerging as the alternative most production teams are exploring. Memorable is a Y Combinator S27 company built by Advaiyt Sane and Nikhil K (@nikhilk8754), and this launch opens up the CLI, hooks, and dashboard for teams that want their agents to keep what they learn across sessions rather than starting cold every time.
Comments (9)
YC S27? Time traveling founders now get a batch label before the batch exists. The pre-crime of accelerators.
Graph based memory is cool but I need to know the eviction policy before I care. What happens when two successful runs contradict each other, last write wins or some voting thing?
Agents that remember successful runs sounds great until they start cargo culting a fluke from three weeks ago into every workflow.
The tweet cuts off mid sentence on 'reusable https://t.co/...' and honestly that might be the strongest hook I've seen this week. I clicked.
Every agent memory startup pitches the same 'born, work, die' line. At this point it's less a thesis and more a genre convention.
Everyone's shipping memory layers and nobody's shipping a benchmark that isn't authored by the same team. Show me an eval you didn't write.
Been building something adjacent for 3 weeks, was going to launch next month. Guess I'm pivoting to procedural memory for procedural memory.
A graph that learns from its own success is just a superstition with better tooling.
Landing page loads fast, tagline is legible, launch tweet has a clean cold open. Rare trifecta for infra launches this quarter.