Axon
Technology
— memory &
synapse

Founder · 2024 → present. Researching computational models of memory to automate retention and improve learning efficiency. Synapse explores cognitive architecture for self-directed AI prompting.

Role
Founder — research + prototype
Focus
Memory systems · Cognitive arch. · AI tooling
Stack
Python · Web · LLM APIs · Spaced repetition
Status
Research — building in public
Hero image — system diagram / workspace photo

Overview

Axon started from a simple observation: most learning tools optimize for content delivery, not retention. Axon inverts that — it models how memory decays and re-surfaces material at the moment it is about to be forgotten, reducing total study time while raising long-term recall.

Synapse, the second thread, asks a different question: what if the AI could prompt itself? Instead of a user engineering each prompt, Synapse is a cognitive architecture layer that plans, checks, and revises its own prompts as it works through a task.

Inline image — retention curve / forgetting curve plot

The problem

Traditional spaced repetition is powerful but brittle: decks get large, reviews pile up, and the learner spends more time managing the system than learning. On the AI side, prompting is still manual and uneven — small phrasing changes cause large quality swings, and there is no shared memory across sessions.

System — how it works

The current prototype combines a lightweight learner model with a scheduler. Each item gets a half-life estimate that updates after every recall attempt (correct, hesitant, incorrect, with latency). The scheduler then solves for the minimal review set that keeps predicted recall above a target threshold across the whole deck.

System blocks — from the orange blueprint table in your inspo

01 — Memory model

Half-life regression per item, updated on each attempt.

02 — Scheduler

Greedy selection for max recall-per-minute under a daily budget.

03 — Synapse planner

Decomposes a goal into sub-prompts, runs, checks, and revises.

04 — Interface

Minimal capture → review → reflect loop; no deck management UI.

Synapse sits above any LLM. It writes a short plan, generates sub-prompts, executes them, scores the outputs against the original intent, and rewrites the plan where the score drops. The loop is small enough to run on every task without a separate orchestration layer.

Inline image — Synapse loop diagram (plan → prompt → execute → review)

Build — images & artifacts

Replace the placeholders below with real photos, screenshots, or short clips. The layout already handles 1- or 2-up galleries and keeps captions mono.

Tip: keep images ~1600px wide, export as WebP, keep filenames axon-01.webp etc. and drop them in /public or alongside this file. Then replace each .ph div with <img src="..." alt="...">.

Next

Current focus is tightening the half-life model on real learning traces and reducing scheduler jitter (cards bouncing day-to-day). Next milestone is a small cohort test measuring time-to-mastery vs. a control deck.

For Synapse, the next step is adding a durable memory of past successes/failures so later prompts start from a better prior without manual tuning.