Engram

Engram is a learning system that installs into a coding agent. Given a topic it builds a first-principles concept map, makes the learner predict, attempt and explain each concept before it is explained, then hands the learner’s verbatim answers to a separate agent that grades them blind: the assessor receives the claim, the rubric, the probe, the production and a confidence the learner picked beforehand, and never sees the lesson. Every verdict is written to disk as a receipt, and nothing moves state without one. One stdlib-only Python file carries the FSRS-4.5 scheduler, the state machine, the statistics and 315 self-checks; it computes every date and stability value itself and contains no network code, which a permanent self-test proves by parsing the file’s own syntax tree. Threshold concepts get interactive HTML explorables built to a seven-clause contract. The same skills and engine run on nine agent platforms, from Claude Code, where it began, to Codex, Antigravity and DeepSeek Harness.

Who built itA single maintainer who works under a handle rather than a name: the GitHub profile carries 38 public repositories, 75 followers, an account opened on 2022-01-03 and a three-word bio, and nothing else. He wrote 189 of the repository’s 204 commits, under two commit names, and the README’s closing section lists four sibling plugins from the same workshop — one of which it credits as the source of Engram’s verification patterns: oracle-driven loops, receipts and re-anchoring.

1. 01

   Fifty-five releases in fifty-three days, under two commit names

   The repository was created on 2026-07-05 and its last commit is dated 2026-08-27: 204 commits in fifty-three days, 162 of them in July and 42 in August. They resolve to four accounts — nagisanzenin with 182, nagisanzeninzz with 7, luanweslley77 with 11 and mertso13 with 4 — while the author’s own commits are signed with two different names, nagisanzenin for 124 and quan-vin for 65, and 180 of the 204 share one Gmail address that a contributor greets in issue #10 as “Hi Quan”. The project published 55 releases and 55 tags. On 2026-07-11, fourteen releases landed between 06:00 and 13:06 — v0.6.0 through v0.6.4 inside forty-one minutes, then v0.7.0 through v1.0.0 over the next four hours. On 2026-07-24, nine releases — v1.3.0 through v1.9.1 plus the 2.0 release candidate — were published inside forty seconds of one another, each titled with that same date. The repairs are as fast as the features: the release titled “what the post-release review caught” follows the release it corrects by thirteen to eighteen minutes, four times over. Of the 204 commits, 144 carry a co-author trailer and 143 of those name a Claude model — Fable 5 on 90, Opus 4.8 on 43, Opus 5 on 6. Around it: 1,439 stars, 105 forks, 5 watchers, 4 open issues, MIT and 2,221 KB, most of it a 783,600-byte scripts/engram.py, which is why GitHub labels the repository Python.

2. 02

   What a receipt contains, and what it takes to earn one

   The evidence that you really recalled something is a JSON object on disk, written at the moment of grading. A receipt carries the topic and the node, the kind of encounter (encode, review, transfer, audit or pretest), the probe that was asked, the learner’s production verbatim, the confidence they stated before any sign of correctness, the grade (recalled, partial or lapsed) and the rating it maps to, one line per distinct misconception, a criterion-by-criterion note quoting the rubric, and the scheduler’s numbers: stability before and after, the interval in days, the retrievability estimate and the next due date. Two fields exist purely as guards. sid, the stash id, rides stash to assessor to receipt so that re-applying a settle file is a no-op instead of double-counting a review; grader holds the assessor’s stable identity, and the specification forbids a model to name its own weights. What makes a receipt count is structural: the assessor never sees the lesson, only the claim, the rubric, the probe and the words, and no state advances without a receipt — a node’s state, fsrs and artifact fields are stripped from anything an external payload supplies. The README offers its own worked example as the argument: a session the tutor thought went well came back 1 recalled, 4 partial and 1 first-retrieval, and the schedule believed the assessor.

3. 03

   The scheduler is code, and a fit that loses is thrown away

   Everything that looks like a date is computed by engram.py and never by the model. The README frames it as the gap a chat cannot close — no memory of you, no test of whether you got it, and no plan for the forgetting that starts the moment you close the terminal — so the plan is FSRS-4.5, with per-node stability and difficulty, a target retention of 0.90 and one fit per learner. refit starts with a single interval multiplier once there are fifty usable reviews, then fits the learner’s own parameters: initial stability at 64 usable reviews, the full weight vector at 400, and a fit that does not beat the current one is refused rather than shipped. The first issue in the repository is a reader’s code reading: next_difficulty mean-reverted toward init_difficulty(4) where FSRS-4.5 reverts toward init_difficulty(3), so the engine had been quietly running FSRS-5’s difficulty rule under a 4.5 label; it was fixed in v0.3.0 and pinned by a fixed-point self-test. The same report caught that a receipt was written after the graph was saved, so a crash could leave an unverified advance; the receipt is now written first, making the worst case a harmless re-review. The first two intervals after encoding are capped so at least three spaced sessions land inside the first month, and a capped queue is ranked by expected retention saved per minute rather than by most overdue.

4. 04

   The grader is graded, and the gold set failed first

   Because the verdict drives mastery, retention, calibration and the schedule itself, the project ships an adversarial gold set of 89 items — fluent-but-empty, terse-but-correct, confident-and-wrong, right-answer-wrong-reason and three analogy-alignment traps added in v1.14 — and reports one number: 0 of 258 blind judgments where the grader awarded more credit than a strict reading of the rubric, earned across three runs on the 86-item version, guarded by a stale-gold check that expires the badge when the set changes. How it arrived at that number is the part worth keeping. v0.7.0 shipped a QWK 0.93 badge, then a post-release reviewer graded the gold set with a deliberately fooled grader; the fooled grader scored higher, 1.000 against 0.990, because five lenient adjudications by the set’s own author credited adjacent facts as partial credit. Correcting them moved agreement from 0.889 to 0.965, and the engine now prints the circularity on every audit: the corrections were prompted by the grader’s own disagreements, so the agreement that follows measures the author’s willingness to concede. One genuine disagreement is deliberately left in, with the reasoning published: an instrument with no disagreement left in it measures nothing. The engine refuses to certify a grader on consistency alone, because a judge can be reproducible and systematically wrong at once.

5. 05

   Explorables built to a contract and registered on the graph

   Interactive HTML is generated for the concepts whose content rewards manipulation — a parameter to drag, a process that unfolds — and never because a learner called themselves a visual learner. Each artifact must satisfy a seven-clause contract: nothing unlocks before a committed guess; at least one guided manipulable model runs the predict, act and explain cycle, with a worked drive first for novices; two free-recall prompts sit inside it with their answers hidden until attempted; decoration is zero and text never runs over motion; the file is self-contained, opens offline and stays under about 120 KB; it ends by asking the reader to rebuild the argument from nothing; and its header records the node, the date and the learner-model inputs it was built from. The file is registered on the graph with artifact set, which validates that it exists, and a receipt stamps whether an artifact existed at grading time — which is what stats.modality compares, explorable-encoded recall against dialogue-only recall, with at least six items per arm. examples/ ships two of them, one generated by the artifact-smith in a real v0.5.0 session and one hand-authored as the reference implementation; both are offered with a hosted link that returns 404 at the collection date, and the only other trace of that publishing route is a zero-byte .nojekyll at the repository root.

6. 06

   Two outside contributors, and five weeks of quiet

   Four accounts have committed here and two of them are the author’s. luanweslley77 wrote eleven commits and five pull requests, several through multiple rounds of review — moving the plugin source out of the directory it extracted into, replacing the instruction file with a versioned marker block, and twice proposing a deletion that would have taken user data with it. mertso13 added Antigravity support and produced agy plugin validate output from version 1.1.4. The sharpest writing is in the bug reports: tyteachestech found that stash add clipped productions at 800 characters before the blind assessor ever saw them, so a thorough answer lost two rubric criteria (fixed in v1.11.2, with the ceiling raised to 2400); DorusKeijzer reported that probes and rubrics routinely disagreed, which became the probe_gap field and a rubric-repair command in v1.10.0; and SK-DEV-AI filed two defects against the OpenCode install with file and line numbers, both fixed in v1.13.2 after the maintainer reproduced the reporter’s counts. The maintainer files against himself as well, and one release is titled “a regression my own fix caused”. The last commit is 2026-08-27. An issue arrived on 2026-09-22 and has no reply, four issues are open, and nothing is archived: five weeks of quiet after eight weeks at that pace is why this record says maintained rather than active.