The DeepVero view
The thesis in one minute
Memorable is building a procedural-memory layer for AI agents: instead of asking a model to rediscover a plan on every run, it captures the successful sequence of tool calls, turns that trace into a reusable procedure, and retrieves the procedure when a similar task returns.
The problem is credible and increasingly important as agents move from demos to repeated production workflows. The open question is not whether agents benefit from reusable workflows—peer-reviewed research says they can—but whether Memorable can become the durable control point rather than a feature absorbed by agent platforms, model vendors, or open-source runtimes.
“Companies will own a durable memory of how their work gets done.”
Memorable founders · Read the original vision on Y Combinator →
01 · Problem
Agents remember facts better than they remember how work succeeded
Most agent-memory systems are optimized to retrieve facts, documents, preferences, or prior conversation snippets. That helps an agent know what happened, but it does not necessarily preserve the ordered actions, tool choices, verification steps, and failure recovery that explain how a task reached a good outcome.
Without procedural memory, a repeated task can still trigger a fresh reasoning loop. The model re-reads context, selects tools again, explores dead ends again, and may choose a different route each time. The economic cost is more than token usage: extra turns add latency, tool/API charges, and more opportunities for execution error.
- Cold runThe agent replans, explores tools, and rediscovers a working path for every repeated task.
- Capture a successful traceA verified run is distilled into reusable procedural steps.
- Memory-assisted runThe agent retrieves the proven path and reasons only about missing or changed steps.
A 2025 ICML paper on Agent Workflow Memory found that induced workflows improved relative success rates by 24.6% on Mind2Web and 51.1% on WebArena while reducing the steps used on successfully completed WebArena tasks. The paper validates the mechanism category, not Memorable specifically.
The pain becomes valuable only under repetition
Memorable is not equally useful for every agent task. It is strongest when work has a recurring structure, a verifiable outcome, and enough volume for saved reasoning to compound. One-off creative or ambiguous tasks offer less reusable procedure and therefore a weaker return.
| Workload | Why the pain repeats | Procedural-memory fit |
|---|---|---|
| Coding agents | Repositories repeat test, debugging, release, and migration patterns | High when outcomes are testable |
| Support and operations | Teams repeatedly navigate the same systems and policy paths | High for stable, governed workflows |
| Browser/computer use | Interfaces require recurring sequences but layouts can drift | Medium–high with robust fallback |
| Research agents | Methods repeat, but sources and conclusions change | Medium; procedure transfers better than answers |
| Novel strategy work | The desired path is ambiguous and changes by problem | Low; fresh reasoning remains central |
02 · Product
Memorable converts successful traces into a navigable workflow graph
The product sits beside an agent harness rather than replacing the model or agent. Its architecture can be understood as a four-stage loop.
- Observe agent tracesCollect successful tool use, decisions, and outcomes.
- Synthesize workflowsConvert raw traces into reusable procedures.
- Assemble the graphConnect procedures, dependencies, and reusable subpaths.
- Retrieve a short pointerGive the next run only the relevant procedure instead of the full history.
- Capture the trace. Record the prompt, ordered tool calls, allowed arguments, and known outcomes from an agent session.
- Synthesize a procedure. Remove dead ends and retain the steps, touched files, checks, and outcome that produced success.
- Assemble the graph. Link procedures through common prefixes and reusable steps so related workflows can be composed.
- Retrieve at run time. Match a new task using exact, lexical, and semantic retrieval, then inject a short procedure pointer into the agent context.
Company-reported Memorable says exact recall takes roughly 60 milliseconds, procedures can live in a local file, gbrain, or Postgres, and only prompts plus allow-listed tool arguments leave the machine for cloud extraction or embeddings when required.
Supported The public gbrain integration documents the actual relay and consent gates, while also warning that the Memorable CLI is closed source and that server-side statelessness and retention behavior remain vendor claims that gbrain cannot audit.
Tool arguments can contain commands, file paths, URLs, and operational identifiers. Local storage and explicit consent reduce exposure, but enterprise adoption will still depend on auditable egress controls, retention guarantees, access policy, and deployment options.
03 · Why now
Agent adoption is moving faster than reliable operational learning
Deloitte's 2026 enterprise survey found that nearly three-quarters of companies planned to deploy agentic AI within two years, yet only 21% reported a mature governance model. Only 25% had moved at least 40% of AI pilots into production. That gap matters: as pilots become production systems, teams need reliability, cost control, auditability, and reuse—not only smarter model outputs.
Procedural memory also benefits from a structural shift in agent economics. Long-running agents repeatedly pay for planning turns and tool execution. Quartermaster's public work argues that turn count and output tokens can matter more to actual cost than raw context-token reduction once prompt caching is considered.
Demand pull
More production agents, repeated workflows, and multi-agent systems increase the number of opportunities to reuse prior execution.
Technology push
Tool traces, verification commands, embeddings, and graph retrieval now make successful procedures observable and retrievable.
Economic pressure
Agent buyers increasingly care about cost per solved task, latency, and pass rate—not only benchmark intelligence.
Governance pressure
A stored and inspectable procedure may be easier to audit than an agent that improvises a new plan on every run.
04 · Opportunity
The opportunity scales with repeated runs, not the headline AI-agent market
Pricing, customer count, and paid usage are not public, so a conventional revenue TAM would create false precision. The more useful bottom-up question is: how much annual execution cost could reusable procedures remove, and what portion could a memory vendor capture?
| Scenario | Target teams | Runs / team | Repeatable share | Avoidable cost / run | Capture | Annual value pool |
|---|---|---|---|---|---|---|
| Conservative | 1,000 | 0.5M | 20% | $0.05 | 10% | $0.5M |
| Base | 5,000 | 2M | 30% | $0.10 | 15% | $45M |
| Upside | 20,000 | 5M | 40% | $0.20 | 20% | $1.6B |
DeepVero estimate Every input in this scenario table is illustrative. It is a sensitivity surface—not a forecast, valuation, disclosed pipeline, or claim about current market spend. The 3,200× spread between conservative and upside cases shows how little a single TAM number would mean today.
The base case requires both scale and monetization power
In the base scenario, 5,000 teams each run two million agent actions annually. If 30% are repeated workflows, if procedural reuse avoids ten cents of model/tool/latency cost per affected run, and if a vendor captures 15% of savings, the resulting annual software value pool is $45 million.
The estimate is most sensitive to workload volume and avoidable cost per run. A team may value reliability more than compute savings, but that value is harder to price without paid retention data. The first validation priorities are therefore production run volume, independently measured cost per solved task, and willingness to pay.
05 · Buyers and go-to-market
The natural buyer owns agent reliability or infrastructure spend
Initial users are likely developers and AI-native teams that can install a CLI and evaluate workflow reuse directly. The economic buyer changes at enterprise scale: platform engineering, AI infrastructure, support operations, or automation leaders will care about success rate, cost, governance, and cross-team reuse.
| Buyer | Job to be done | Proof required |
|---|---|---|
| Agent developer | Stop rediscovering known implementation paths | Fast install, useful recall, no workflow breakage |
| AI platform team | Improve cost and reliability across many agents | Harness coverage, controls, observable ROI |
| Operations leader | Make automated workflows consistent and auditable | Completion rate, exception handling, policy enforcement |
| Security / governance | Control what traces leave and who can reuse procedures | Audit logs, data residency, deletion, access controls |
Company-reported Memorable is publicly recruiting design partners running agents at volume and offers custom harness integrations. No public pricing page, paid-customer count, revenue, or retention data was found as of September 23, 2026.
DeepVero estimate A credible long-term model is infrastructure SaaS priced by organization, stored procedures, or agent runs, with an enterprise tier for self-hosting and governance. This is an inference, not announced pricing.
06 · Competition and moat
Memorable must own the procedure layer before platforms bundle it
The competitive set is broader than products labeled “memory.” Buyers can approximate the outcome with hand-authored skills, workflow engines, retrieval systems, observability traces, or memory built into their agent platform.
| Alternative | Strength | Gap Memorable targets |
|---|---|---|
| Vector / text memory | Strong retrieval ecosystem and familiar RAG patterns | Returns knowledge chunks rather than verified action sequences |
| Hand-authored skills and playbooks | Explicit, reviewable, and easy to version | Expensive to maintain; learning is not automatic |
| Deterministic workflow engines | Reliable for known, stable processes | Rigid when tools, context, or task variants change |
| Native agent-platform memory | Bundled distribution and deep runtime access | Can be vendor-specific and less portable across harnesses |
| Trace / observability platforms | Rich data on what agents did | Usually diagnose runs rather than automatically reuse the successful path |
The strongest moat would be organizational learning, not the graph itself
A graph representation and hybrid retrieval are reproducible technical ideas. A more durable advantage could emerge from a growing corpus of outcome-linked procedures, deep integrations into agent runtimes, trusted governance, and compounding improvement as procedures are reused and revised across a company.
That moat is not established yet. Early teams have little proprietary procedure data, and platform vendors already see the same traces. Memorable must become the system of record for how agent work succeeds—not merely a convenient retrieval plugin.
07 · Evidence and traction
The early numbers support continued testing, not a broad performance claim
Quartermaster
Memorable reports tool calls falling from five to three across three replications, with pass rate at 91% with memory versus 80% without. The published scope is replay of a known task family.
gbrain
The company site reports turns per task falling from 16 to 13 across 454 passing runs on a three-bug coding fixture.
gstack
A learned procedure is reported as 293 tokens versus 15,593 tokens for an investigation skill—a 98% context reduction.
OpenHome
The company site reports task completion improving from 70% to 100% for a design partner.
The YC launch post states 60% fewer tool calls and 40% lower latency on “QM's benchmarks,” while the current product site highlights a 40% tool-call reduction for Quartermaster. The difference may reflect distinct fixtures or revisions, but the public materials do not fully reconcile it. Context reduction also should not be treated as equivalent to cost reduction under prompt caching.
Supported Public integration documentation confirms that Memorable is wired into gbrain and Quartermaster and describes the data path and safeguards. That is stronger than an unsupported product demo, but it is not the same as independent, powered evaluation across diverse customers and task distributions.
08 · Risks
Five failure modes could compress the opportunity
- Bundling: model and agent-platform vendors can add native procedural memory and distribute it to an installed base.
- Workflow drift: stored procedures can become stale when tools, permissions, interfaces, or business rules change.
- Trust and security: commands and paths can be sensitive; a closed-source client increases the enterprise proof burden.
- Weak transfer: performance on repeated task families may not generalize to novel tasks or materially changed environments.
- Economics: if model inference becomes cheaper faster than agent workloads grow, compute savings alone may not support strong pricing.
09 · Investment thesis
A credible wedge with an unproven control point
What to believe: procedural memory is likely to become a meaningful component of production agent stacks. Memorable has a clear product thesis, working integrations, and early evidence aligned with academic results. The company is targeting an operational pain that should intensify as agent volume increases.
What remains unproven: whether customers will pay for a standalone layer, whether savings persist across diverse production tasks, and whether accumulated procedures create defensibility before native platforms catch up.
Signals that would strengthen the thesis
- Independent reproduction of lower cost per solved task across multiple harnesses and task families
- Paid design-partner conversion and strong six- or twelve-month retention
- Growing procedure reuse per organization without rising stale-procedure failures
- Enterprise-grade controls for audit, deletion, residency, and role-based sharing
- Evidence that cross-agent procedure graphs improve outcomes beyond single-agent replay
Signals that would weaken the thesis
- Benefits disappear on novel or evolving tasks
- Customers retain the product only while receiving custom integration support
- Native model or agent platforms match the capability with lower deployment friction
- Security review blocks trace capture in the highest-value enterprise workflows
- Token or tool-call reductions fail to improve total cost, latency, or completion rate
This profile is an analytical company teardown, not investment advice. Memorable is private, financial information is limited, and scenario values are illustrative.