Tortoise by Premise Labs
GX Fellowship Memo · working draft · 2026-08-27

Tortoise.

Memory for agents to remember why, not just what.
Prepared for
Jesse · Adam · Nak
Date
August 2026
Author
Daniel Ospina
Company
Premise Labs (tortoise)
Cover Letter

For Jesse, Adam & Nak —

Tortoise provides memory for agents that computes how knowledge changes. Current memory is text: Markdown files or knowledge graphs that store semantics. Facts pile up; every update makes the agent rediscover the logical connections and dependencies; inconsistencies multiply and drift makes hard-earned insights unrecognizable. That's why agents can't yet run without supervision; they keep wasting tokens reinventing the wheel or undoing previous work.

Instead of storing memory as text, Tortoise builds a logical, probabilistic world model:state (what is), context (who and when), and epistemics (the reasons and beliefs behind decisions). New data doesn't trigger rewrites. It connects to the model; confidence is quantified structurally, from sources and the web of arguments that support, negate or mitigate, and a deterministic algorithm calculates how new findings affect every node in the graph. Decisions stay auditable; drift is minimized. Agents reason with hundreds of variables instead of losing the plot past a few. Multiple agents collaborate on a shared world model and build on previous learning.

Both individuals and companies are already trying memory solutions, and they're dissatisfied. My tweet explaining the problem with traditional solutions and how we think about Tortoise reached 1M+ views, 850+ likes, and 17 inbound leads (and counting); 15 of 17 confirmed the dissatisfaction and willingness to try. The leads are builders: devs, PMs, and tech-industry veterans, heavily vibe-coding. 60% want Tortoise for personal knowledge management, 40% to power an agentic application, and one also wants to introduce it to the rest of his company.

The economics: at measured per-session costs and typical usage, a regular user is ~$2/mo all-in (storage ~$0.1, LLM ~$1.8) against a $25 Pro / $149 Team price, deeply margin-positive. Heavy usage runs $12–37/mo and is usage-metered; extraction (internal to hosted, being optimized) is the calibration lever, and the business model exists for exactly this (M3 ≈ $3K MRR, M12 ≈ $99K MRR, Y5 ≈ $101M ARR). Default alive by month ten: gross burn to default alive ≈ $640K, with a 50% buffer, a total ask of ≈ $960K.

We're entering this market bottom-up with the segment we already have demand in: vibecoders and indie hackers. They onboard in five minutes and get value from the first session by making a decision on the graph (which vitamins to buy, which screen). As AI's early adopters, they happily wire their own workflows and integrations, and many contribute in open source. They help us progress up the food chain by acting as recommenders, and help us scale horizontally by building vertical apps.

I started my career as a chef in the best restaurants in the world, and found they struggled to remain innovative, lacking systems to collaborate on and manage their knowledge. That sent me on a decade-long quest to understand why smart groups fail to think together: experimental psychology at Oxford, organization design consulting (BCG, Oxford visiting lecturer, etc.), collective-intelligence tools for DAOs, and a previous failed startup that gathered community data into an AI hive mind. Then I started using agents and stumbled upon the same problem, now more pressing than ever. I paused everything else and jumped head-on.

M12, that's a ~$1.2M ARR business; in five years, ~$101M ARR at ~97k customers, profitable along the way, on Gateway X's own thesis: build on your own terms, don't pass the valuation on.

— Daniel Ospina · Founder, Premise Labs · August 2026

Contents

Where we are · August 2026
Alpha
in production, powers every decision in my own agent stack (this memo was written on it)
17
inbound leads from thesis tweet (and counting) · 15/17 confirmed dissatisfaction + willingness to try
1M+
views on the thesis tweet · ~850 likes
1–2 wks
to private beta · waiting cohort of testers
Question 1

What is it?

The bottleneck for intelligence isn't compute; it's learning. Agents start forgetting after a few thousand tokens, and today's memory systems are text: Markdown files or knowledge graphs that store semantics. Facts pile up; each edit makes the agent rediscover the logical connections; drift soon makes your knowledge unrecognizable, and hard-earned insights are lost.

Tortoise is a memory system for agents that enables them not just to store and recall facts, but to understand the logic behind decisions, comprehend how each component fits into a larger whole, and learn over cycles without drift.

Category.

Memory infrastructure for agents. Every relationship expresses how one belief supports, contradicts or mitigates another; when new evidence arrives, an algorithm propagates confidence through the graph, updating the world model without rewriting files or recomputing from scratch. Memory becomes an active reasoning system, not passive storage.

Form factor.

A foundational service: use it solo with your agents, as infrastructure to power applications, or together with your team. Hosted or self-hosted.

Question 2

Customer + edges

Who.

  • AI builders: devs, product managers and tech veterans who are building with agents ("vibe coding"), in entrepreneurial communities. The 17 inbound leads fit this profile exactly: 60% want Tortoise for personal knowledge management, 40% to power an agentic application (two of them building investment applications), and one wants to use it personally and bring it to his company.
  • Later: teams (pending permission management and automated curation) and enterprises deploying autonomous workflows (pending compliance, audit).

Why this segment.

Indie hackers and vibe coders are AI's neophiles and early adopters, and they're capable of building workflows and integrations themselves; many are open-source enthusiasts. Set up right, the platform gives them a contribution loop (skills and integrations marketplace). From there, the roles compound: self-users recommend to companies and build applications on Tortoise.

Knowledge edge.

Memory is a learning problem, not a retrieval problem
Competitors help agents find information. The harder problem is computing how knowledge changes. Knowledge graphs are only semantic; when something changes, an agent must manually update everything, making them fragile and messy. We made the graph computational by adding quantitative, epistemic relationships and an algorithm for belief propagation.
A decade on why smart groups fail to think together
Michelin chef working in R&D → experimental psychology (Oxford) & London University Philosophy → Organization Design consulting & cybernetics → collective intelligence (DAO governance, decision markets) → Community Hivemind startup → agent memory. Current product is the fruit of years of R&D, including a provisional patent.

Distribution edge.

Founder-led content
Thesis tweet with 1M+ views and ~850 likes, leading to inbound leads. Soft skills I learned over a decade doing thought leadership content (article in Harvard Business Review, 2 TEDx talks, etc.).
Channel plan
Content-first (Twitter and LinkedIn to start, then video: YouTube, Instagram, and Telegram). Founder-led, amplified with ads and referrals.
Question 3

White space

Two reasons this doesn't already exist:

01 · The technology didn't exist, until now
Digesting messy streams into a computational world model only became automatable on current models, at current prices.
02 · The need was retrieval, not intelligence
Existing memory systems are retrieval systems: storage, not computation. They sell "find it faster"; the gap for agents is learning. Same product category but totally different design assumptions.

Alternatives and why they fail this customer

AlternativeWhy it fails
Markdown files + retrieval (the default)Summaries edit the source data: compression, drift. Knowledge gets quickly outdated because agents are lazy and miss some of the logical connections and dependencies between docs.
Mem0 / Zep / Honcho / etcRetrieval systems with a graph. No epistemic layer: unreliable when answering "why do we believe this?" and can’t propagate how new evidence changes beliefs without drift or significant token costs.
Flagship AI LabsCompanies don’t trust them with their data. The Labs’ incentives are for increasing usage or tokens (e.g., editing markdown files), not for providing higher intelligence with fewer tokens.
Company BrainsThey’re spread thin trying to do everything (harnesses + integrations + memory, etc.) in an era when customers are becoming more accustomed to customization, and agents can complete an integration for virtually no cost. We can integrate with them later on.

The gap over time

Models keep improving and retraining will become more common, but many kinds of knowledge (business strategy, R&D, investment decisions) need to evolve continuously while staying transparent and deterministic. That belongs in memory, not inside model weights.

Question 4

Why now, how big

Why now

  • Cost crossed the line. Running agents reliably (extraction, embedding, reasoning loops) is now cheap enough that the economics work: measured per-session extraction costs ~$0.03 at the median. 6 months ago this was 20x more expensive.
  • Capability crossed the line. Models are now intelligence-wise, capable of doing far more. That's exactly why memory became the bottleneck: the agents that forget the trade-offs, concerns, and arguments behind decisions burn tokens reinventing the wheel or advancing incoherent initiatives. Compute is no longer the constraint; learning is.
  • The category just got named. Customers are actively searching.

How big: the whole model

~97,000 customers at ~$100/mo (blended) can make us $100M ARR, while there are 250,000+ agent-building developers and teams today, and the agent economy is compounding.

HorizonCustomersMRR
M3 (demo day)300 Solo + 1 Team~$3K
M121.4K Solo + 1.1K Pro + 384 Team~$99K (~$115K w/ 2 enterprise)
Y5~97K customers~$101M ARR (agent-count tailwind could 10×)

Assumptions: 10% of customers buy overage; 8% Team with 30-day cycle; CAC $75 Solo / $400 Team; 2 enterprise deals Y1, 8+ Y2 (9–12mo cycle). Top-down sanity: TAM $28B memory & orchestration (Mordor 2030), SAM $5.7B, SOM $114–285M.

Question 5

Show, don't tell

Demo video

youtu.be/Zylw3BB8NoQ

What the machine does

  • Connects your data (feeds, docs, conversations) into a world model: state, context, epistemics.
  • Propagates confidence deterministically when new evidence arrives; no rewrites, no drift.
  • Ranks options in a decision (Tortoise Decide): agents usually get lost past a few variables; here they can reason with hundreds or even a million.
  • Leaves an audit trail: every decision is provenance-bound, showing which source, which reasons, which confidence.

Planned fast path (design, not shipped)

Ship a prompt with ~10 example decisions a user can test in the first session: which vitamins to buy while scrolling Amazon, which monitor, or a real startup decision if they have one. The decision → analysis output is the aha moment.

Question 6

Demand

B2C → B2B: the builders who use it for themselves and to power their applications.

6a · Users: builders, on record

Memory is a real challenge. Played around with OKF and Graphs so far. Currently designing a .md based graph… I'm wondering if Tortoise could be a good fit.

Geoff · cofounder, getmothy.ai (company details unverified) · Aug 2026, DM

Graph surfaced a note vector missed in 72% of them, +2 notes on average. Your drift point still beats us. How does tortoise keep the "why" when the same fact comes back with a different reason?

Tony · builder who ran the vector-vs-graph A/B · Aug 19 2026, DM

Most agent memory systems are still pretty basic. Markdown files with retrieval layered on top… Knowledge graphs are designed to represent what is known, not to continuously learn and update beliefs.

Mehrdad · arscontexta.org (memory infrastructure for Claude Code) · Aug 2026, DM

6b · End customers: through builder apps

The end-customer wedge is the 40% building agentic applications: when those apps commercialize, Tortoise is the infrastructure underneath; the non-technical segments they serve become reachable without us building vertical apps. Two of the 17 leads are building investment applications; one lead also wants to bring Tortoise to the rest of his company (a Team-tier signal).

Channels

Founder-led content on Twitter and LinkedIn: the thesis tweet is the proof (1M+ views, 17 leads); strategy is content-first amplified with ads and later on referrals. Scaling via paid UGC (influencers).

Question 7

Economics

Pricing: feature tiers by use case, usage on top

TierPrice/moWrite ops incl.NodesOverage
free$010K10K
solo: use it yourself$910K25K$5 / 10K write ops
pro: build with it$2550K100K$5 / 10K write ops
team: use it together$149200K600K$5 / 10K write ops

Tiers are feature tiers by use case: solo (personal knowledge management, the 60% of our leads), pro (power an agentic application, the 40%), team (company-wide); each with included usage to get going, usage metered on top.

The margin story

A typical user costs ~$2/mo to serve (storage ~$0.1, LLM ~$1.8 at measured per-session costs), against $9, $25 and $149 tiers: ~80% gross margin before usage. Heavy usage is metered on top (usage tokens, extraction weighted), so power users pay their way as the product does more of the work.

Break-even

Operating break-even comes at roughly 1,600 customers, where gross margin covers the future team. Projected default alive by month ten, on a gross burn of ≈ $640K, with a 50% buffer: a total ask of ≈ $960K, comfortably within Gateway X's check.

Question 7b

Compounding

01 · Usage tracks the customer's own agent scale
Revenue is usage-metered on top of tiers. As a customer's agents do more work, write ops grow; we compound with their adoption. The agent economy's growth is our growth curve.
02 · Shared world models get stickier and more valuable
The graph is a living model of decisions and their reasons. Every decision a team files makes the model more useful for the next agent and the next hire: a data moat that grows with use.
03 · The skills & integrations marketplace
Many builders are open-source enthusiasts. Give them a marketplace to share, and they help us extend the platform. A community that grows with adoption.
Question 8

Wedge + resolution

The entry point is the builder segment we already have demand in (17 inbound leads, two building investment applications), entered through founder-led content and MCP/library distribution, at the Solo/Pro price, cheap enough to explore, which is the point.

Time to value, the magic moment: Tortoise Decide
Same session: install, connect your data, make a decision with options ranked by propagated confidence. Pick an example decision (e.g., which consumer electronics to buy, or a business one), connect the evidence, watch confidence propagate and options rank with an audit trail. Value from the first session, and data ingestion configured for the aftermath.
Why this segment, why now
Indie hackers and vibe coders are AI's neophiles and early adopters, and they can wire workflows and integrations themselves. Many are open-source enthusiasts. The roles then compound: self users who recommend, who build applications on Tortoise, who advise companies on what to implement. We already have demand here.

How the model resolves

Horizontal infrastructure so we can focus and double down on our USP of a technical edge. Builders customize or ship the vertical apps.

Question 9

Meant to build

I started my career as a chef, working in some of the best restaurants in the world, until I realized my culinary heroes were struggling to remain innovative because they had no good way to manage tacit knowledge. I then spent a decade on the question: why do intelligent people, together, get dumber? From a residency in experimental psychology at Oxford, to considering an MBA but ending up instead collaborating with one of the professors and facilitating workshops for their executive education programs, training BCG consultants, and then moving to work on DAO governance and launching a Web3 venture studio focused on collaboration tools. Seeing the dysfunctions of aggregating knowledge in forums and chat groups, we launched a startup to organize that data into a Hivemind to power a chatbot. Generated $100k+ of revenue, but the Web3 market didn't grow as we hoped.

A year later, building agents, the same problem returned: LLMs forget halfway through a session, and every memory system is text with retrieval layered on top. I suddenly realized I had all the pieces to solve this problem, and paused everything else to focus on it.

Team by archetype

ArchetypeWhoStatus
Founder / product / salesDaniel OspinaSolo founder; built MVP + validated algorithm; owns the funnel
ResearchConnorAdvisory stake (2%); co-developed the logical-chains ontology
CTO & CMOCurrent gapsKey hires post-raise

Iteration receipts

My original go-to-market was going to be a couple of example applications that memory powers better, including a bot for founder-led content that uses the logic of the memory. Then the thesis tweet went insanely viral and generated inbound leads. The demand was for the memory system itself, so I pivoted the go-to-market away from the example applications.

In a conversation with a potential customer, they showed me the detailed ontology they had developed within their enterprise. And another solo dev also mentioned how controlling the ontology is important for them. I quickly moved ontology expansion packs as a roadmap priority.

Self-reflection

TraitReceipt
Compass points northI won’t sell memory I wouldn’t trust in my own stack; Tortoise powers every decision I make, including this memo.
Action-orientedWorking alpha in 3 weeks; provisional patent filed; thesis tweet 1M+ views; daily calls with customers.
Admits unknownsExtraction cost poorly measured (improving now); wedge is decided but unproven.
Stake-honestTeam gaps named on the team table.
Question 10

Why Gateway X

I went through Techstars with a previous startup and learned the hard way how valuable mentors are when they have skin in the game. I want partners who prioritize telling me the truth over being nice and maintaining the relationship. The investment size here is attractive because it eliminates distractions: enough to reach default alive and never raise again.

And I strongly resonate with the thesis of building profitable businesses:maximizing market share is not necessarily the best path to profitability.