Hi HN, entityone here.
Short version of how this exists: I was pair-programming with an LLM and got tired of re-explaining the same context every morning. So I spent about ten days building it a memory. Atoms (named, versioned strings) hashed into a SHA-256 Merkle tree, an RFC 6962 consistency proof so I could verify the tree evolved honestly, and a Prediction-by-Partial-Matching model over the recall sequence so it pre-fetches the next atom before I ask. Sub-millisecond recall. I was very pleased with myself.
Then I hit the problem that is, in hindsight, extremely funny: I had built a memory system and had nothing to remember. A proof of honesty over an empty tree proves nothing. You cannot evaluate a memory substrate on a toy. It needs a real project with real bugs, real corrections, real state that changes under you and lies to you.
So I built one. The "test harness" turned into an actual SaaS — billing, provisioning, DigitalOcean orchestration, the works — written almost entirely with the same AI, using the ten-day memory engine as its long-term memory the entire time. Roughly eight months. The product became the experiment. The experiment became the product. I am aware this is backwards.
A few things I did not expect: corrections become durable (the first time I told it "ask, don't guess," it stored that as a procedure and it's loaded on every session since); bug shapes are retrievable (a three-hour crash-loop debug is now one atom it reaches for first); and the compounding is real but slow — marginal for a few weeks, then around 200–300 atoms it flips and sessions start with it already knowing what I'd have re-explained.
It's MCP-native (Claude, Cursor, Cline connect with no SDK), there's a REST API, and a 3D Merkle-tree visualiser because I wanted to watch the proofs.
Honest limitations: the 64% Markov next-atom hit rate is measured on my own sessions — yours will differ, and I'd like more external numbers. It's single-tenant by design (your own substrate), so no shared-corpus magic across customers. And the discipline is on the human: it only knows what you bother to make it remember.
Repo: github.com/wjm2202/Parametric-Memory. Happy to get torn apart in the comments — the Merkle and PPM internals are the fun part to argue about.
Months of working with this second brain has proven to me this needs to be shared with the world
[dead]
Hi HN, entityone here. Short version of how this exists: I was pair-programming with an LLM and got tired of re-explaining the same context every morning. So I spent about ten days building it a memory. Atoms (named, versioned strings) hashed into a SHA-256 Merkle tree, an RFC 6962 consistency proof so I could verify the tree evolved honestly, and a Prediction-by-Partial-Matching model over the recall sequence so it pre-fetches the next atom before I ask. Sub-millisecond recall. I was very pleased with myself. Then I hit the problem that is, in hindsight, extremely funny: I had built a memory system and had nothing to remember. A proof of honesty over an empty tree proves nothing. You cannot evaluate a memory substrate on a toy. It needs a real project with real bugs, real corrections, real state that changes under you and lies to you. So I built one. The "test harness" turned into an actual SaaS — billing, provisioning, DigitalOcean orchestration, the works — written almost entirely with the same AI, using the ten-day memory engine as its long-term memory the entire time. Roughly eight months. The product became the experiment. The experiment became the product. I am aware this is backwards. A few things I did not expect: corrections become durable (the first time I told it "ask, don't guess," it stored that as a procedure and it's loaded on every session since); bug shapes are retrievable (a three-hour crash-loop debug is now one atom it reaches for first); and the compounding is real but slow — marginal for a few weeks, then around 200–300 atoms it flips and sessions start with it already knowing what I'd have re-explained. It's MCP-native (Claude, Cursor, Cline connect with no SDK), there's a REST API, and a 3D Merkle-tree visualiser because I wanted to watch the proofs. Honest limitations: the 64% Markov next-atom hit rate is measured on my own sessions — yours will differ, and I'd like more external numbers. It's single-tenant by design (your own substrate), so no shared-corpus magic across customers. And the discipline is on the human: it only knows what you bother to make it remember. Repo: github.com/wjm2202/Parametric-Memory. Happy to get torn apart in the comments — the Merkle and PPM internals are the fun part to argue about.