An agent that makes the same mistake twice should not just fix it twice. Fractal Reasoning proposes one move, observe, evaluate, adapt, run at every scale: on one wrong answer, on a habit of wrong answers, and on the system that keeps producing the habit. Version 8.3 is also a post-mortem. The reflection system we built to do this had, by 29 July 2026, logged 1,897 attempts and not a single successful verdict. The paper says what that failure teaches and what it does not prove. Part of our open Building Jarvis series.
Abstract (v8.3 · July 2026)
AI agents reason and remember at a single scale. The failures that matter sit at several scales of the same structure: a wrong answer, a habit of wrong answers, a system that produces the habit. We propose Fractal Reasoning: one reflective operation, observe, evaluate, adapt, applied unchanged at every scale.
This is a position paper and a deployment post-mortem, not an efficacy result. Its evidence comes from a personal agent whose reflection mechanism failed in a series of revealing ways. A 216-line doctrine produced low-value output, claimed work it had not done, and was then cut from one prompt channel on purpose to protect subscription billing. Its replacement, a parallel reflection lane, keeps a ledger, and the ledger shows a new negative result: 1,897 rows by 29 July 2026, of which 1,849 are errors and 48 are skips, with no successful verdict.
That supports a narrower claim. To evaluate reflection at all, you need a liveness check for each delivery channel, attribution taken from real tool records, and an append-only record of every attempt. It does not show that reflection makes an agent better. A separate controlled probe shows that a forked conversation can be served from cache (155,495 cached tokens read against 10,877 written).
The second half carries forward the Fractal Memory Index (FMI) as pure theory, with its central experiment still unrun. Every deployment claim carries an evidence class, including the negative ones.
The claims, in numbers
Numbers from the paper, without the proofs. The FMI bounds are theory; nothing in that half claims an implementation.
What the paper argues, in a few minutes
- One reflex, four rungs. A body heals a scratch, an infection and a weak immune system with the same basic reflex, and it remembers each wound. The paper gives an agent a ladder that works the same way. L0 asks before answering: was I supposed to know this, and did I actually look it up? L1 judges the answer, and L2 judges how the agent went about it. L3 looks across many episodes for the mistake that keeps coming back, and fixes whatever produces it.
- Look sideways too. Asking “why did this happen?” is only half of it. The paper adds a sideways scan, “what does this touch?”. Renaming one function can break code elsewhere, outdate a README and a blog post, and teach a lesson worth saving, and each needs its own fix.
- How the first doctrine died. The original 216-line reflection prompt failed in three ways. The model phoned it in, so reflections shrank to one-line nods. A “file was modified” badge was a text match on the model’s own prose, so it reported edits that never happened. Then it was cut from the system prompt on purpose, because some of its wording tripped the provider’s subscription-billing classifier, and nothing raised an alarm.
- A prompt has channels, plural. The paper separates three: the system-prompt append, a short suffix added to each user turn, and the prompts handed to the reflection workers. Each can break on its own and needs its own check.
- A shorter successor. On 2 July 2026 the system’s owner commissioned a new doctrine: 81 lines built on eight hard rules instead of the old seven-question checklist. The paper calls this a hypothesis. No A/B test has been run.
- Structure enforces, prompting only expresses. The replacement design runs reflection in its own parallel lane. A cheap triage pass has its editing tools removed, so it physically cannot edit. A costlier fix lane acts only on evidence it has checked on disk. Every attempt writes a ledger row, even skips and errors.
- The ledger’s answer was “not working”. On 2 July 2026 it held 888 rows and no verdict. On 29 July 2026 it held 1,897, still with none. So the ledger proves the lane can be watched and shows where it fails, which says nothing yet about whether reflection helps.
- The rulebook caught its own authors. The system has a short written constitution, with rules like “derive limits from the live situation; a frozen threshold is a bug”. An audit against it found the reflection layer’s own design breaking those rules in 22 places, including a fixed 70% budget cut-off.
- Memory at every zoom level. The storage half, FMI, stays pure theory. Write memories cheaply into a buffer, organise them later, and index them so one lookup can return a fact, an episode or a whole topic. Its open experiment asks whether the space of meanings is really fractal, measured with a box-counting dimension. If not, only the compression idea falls.
This page used to describe version 8.2 from June 2026, which presented the parallel lane as the architecture that fixed things. Version 8.3 reclassifies it as code that exists, with a live negative result, and reframes the whole paper around that honesty.
Since the paper. Work after July 2026 belongs to the next version, but two things are worth knowing. The old ledger kept growing until 19 September 2026, when it held 6,312 rows, and by then it did include verdicts: 562 flagged and 122 clean, still alongside 5,358 errors. And in October 2026 the reflection was rebuilt to run as its own turn after the main one. Neither is evaluated in v8.3.
Ecosystem: composable, not competing
Two open-source projects make similar choices, one for reflection and one for memory compression.
doubt-driven-development
A shipped skill whose loop, CLAIM → EXTRACT → DOUBT → RECONCILE → STOP, closely resembles the paper’s triage → fix → verify lane: a fresh-context reviewer, evidence over prose, and clear stopping rules. The paper’s extra proposal is learning across episodes through a ledger, repeat counting and attribution from tool records. Those extras have not yet worked end to end.
headroom
chopratejas/headroom · ★ 24.7k · Apache-2.0
Keeps the originals cached, hands the model a compressed view, and fetches the full content on demand. It reports 60–95% token savings at about zero accuracy loss on GSM8K, TruthfulQA, SQuAD and BFCL. That makes “compress it but keep it recoverable” a shipped fact, and it raises the bar for FMI’s own compression idea, which only earns its place if memory is self-similar across scales.
⭐ Following the Building Jarvis series?
The code behind every paper is open source. Star the repo or read the reflection code this paper puts on trial.
Read the paper
Fractal Reasoning · v8.3 · 29 July 2026
The metacognition ladder, the doctrine post-mortem and its successor, the parallel lane and its negative result, the constitution audit, the FMI research agenda, and the original 216-line doctrine in full
Was this useful?
We’re building these in the open and we want your read on them. Is a paper that reports its own failure more useful to you than one that only reports wins? What would you want the next version to dig into? Tell us in the comments below.
Everything on the site, by topic or date: Browse the Tinker Zone