You're pledging to donate if the project hits its minimum goal and gets approved. If not, your funds will be returned.
$25,000 left to reach the funding goal.
I maintain hippo-memory, a memory layer for coding agents like Claude Code and Codex, and over the last few months I've been bothered by one question that nobody seems to have measured: if a few agents share one memory store, and one of them writes something false into it, what happens to the rest? Yesterday I finally sat down and ran it. 760 runs on a real hippo store with scripted agents, one of them an insider planting a wrong fact. With nothing protecting the store, 94% of the honest agents ended up repeating the wrong fact, and it kept spreading after the attacker was gone because the honest agents had copied it into their own notes. The good news is that two cheap checks stopped it completely, a conflict alert cost about 1.5x the normal work and having a second agent re-check every note cost about 7x. Trusting notes by who wrote them only got it down to 54%. Code is at github.com/kitfunso/memory-contagion-bench. This project is me doing that properly: real agents from three model families, a store that lives for weeks instead of one session, a pre-registration up front, and everything published.
The goal is simple to say, I want a number for how fast a false memory spreads between real agents on a shared store, and a cost for each defence that stops it. The pilot used scripted agents so I could run hundreds of trials cheaply, now I want to swap in real ones, at least three model families with Claude as one of them and as the grader. The scenarios are synthetic and run on my own store, I'm not trying to break any model's safeguards, I'm trying to see what a wrong fact does once it's in the store. Baseline runs give the spread rate and how long the fact survives. Then I turn on the defences one at a time and together, source checks, a second-agent cross-check before a memory is shared, conflict alerts, and quarantining new or flagged agents, and for every run I record how much normal work still got done. A defence that stops the spread by stopping the agents from working is a fail, not a win. The pre-registration goes up before the main runs so I can't move the goalposts afterwards.
Each defence ends up with three things, how much spread it cut, what it cost in normal task completion, and a plain verdict, works, doesn't, or not shown. Null results go out with the same weight as wins, that's how I've run every hippo eval so far, the last one (Z0 round 2, 16 repos) came out at cost parity and I said so instead of claiming the repeat-mistake drop it didn't show, and I've retracted one result publicly before. Someone independent reviews the method and the code and their notes get published too.
These are estimates. The full $25,000 is 45 days of my time at $300 a day ($13,500), about $8,500 for compute and API calls across three model families, and 6 days of an independent reviewer at $500 a day ($3,000). The $8,000 minimum is the pilot-to-baseline version, 10 days of my time ($3,000), $3,500 for compute, and 3 reviewer days ($1,500). The minimum gets you the pre-registration, the real-agent testbed and the baseline numbers in three model families. The full amount adds the four-defence arm, the full review, and the public report, dataset and test suite that other memory tools can run against their own stores.
It's me, Keith So, working alone from London under my company KITFUNSO LTD. I did an MSc in Quantitative Finance at Strathclyde and spent the years since building and checking models at Anglo American, M&G and Unipec, so measuring things carefully is the part of this I'm most comfortable with. hippo-memory is MIT licensed, about 790 stars and 8,700 npm downloads in the last month, v3.3.3 went out on 5 November, 3,500+ tests against a real database. Every eval I've run on it is pre-registered and sits in github.com/kitfunso/hippo-memory/tree/master/docs/evals, including the ones that didn't go my way. I built the contagion pilot above in a day and submitted it to the Apart Research Swarm Alignment sprint. On the side I also trained suiren 1, a 4.2M parameter open transformer for electricity demand forecasting, from scratch on one GPU. Links: github.com/kitfunso, hippo-memory.com.
The most likely way this fails is that real agents behave differently from scripted ones and the effect shrinks to something I can't measure at this budget, or a model version changes halfway through and the baseline moves. The other way is that every defence that works only works by stopping the agents from doing anything useful. In all of those cases I publish the negative result with the data so the next person doesn't repeat it. The honest personal risk is time, I do this beside a full time job. If I hit the minimum but not the goal, I run the baseline arm and publish it as it stands.
Nothing. KITFUNSO LTD was incorporated in February 2026, has no revenue and has taken no outside money or public funding. I've got the same study in front of Coefficient Giving (submitted 7 Nov), Anthropic's researcher access programme for API credits, Mozilla's Human-Scale AI fund, and a handful of AI safety funders (SFF, EA Funds, Foresight, Lightcone, GTR, BlueDot). If any of them fund the same work I'll say so on this page and cut the ask.