Memory-Assisted Shaping is a behavioral protocol for ChatGPT Projects that keeps one idea structurally coherent across a long conversation. It gives GPT a small, well-defined continuity layer — explicit operating states, minimal memory signals, and a strict separation between shaping and final artifact production — so that a complex idea can be worked through without the conversation collapsing into noise or the final output carrying process residue.Documentation Index
Fetch the complete documentation index at: https://mintlify.com/xxyoudeadpunkxx/memory-assisted-shaping/llms.txt
Use this file to discover all available pages before exploring further.
Memory-Assisted Shaping is a behavioral protocol, not a chatbot or SaaS product. It is a set of files —
protocol.md, note-io.md, and notes.py — that you add to a ChatGPT Project. GPT reads them at session entry and follows the operating contract they define. There is no server, no account, and no external service involved.The drift problem
Long ChatGPT conversations drift. The problem is not that GPT forgets facts — it is that the conversation accumulates process noise that corrupts structure over time. Specific failure modes:- Examples become decisions. A casual example offered to clarify a point gets picked up and treated as a confirmed design choice in later turns.
- Discarded paths return. A direction you explicitly ruled out re-emerges as a suggestion because nothing in the session marks it as closed.
- Open questions disappear. A real unresolved decision — one that actually changes the artifact — gets smoothed over by momentum and never gets answered.
- Final outputs carry process residue. The artifact arrives with shaping commentary, half-considered alternatives, and reasoning traces baked in, because the conversation never separated working notes from finished output.
What Memory-Assisted Shaping does
The protocol applies a three-layer solution that keeps GPT’s behavior tractable across an entire shaping session. Layer 1 — Explicit operating states. GPT operates in exactly one of three states at any moment:READING_ALIGNMENT (understanding the input and identifying real gates), SHAPING (consolidating form, surfacing and closing gates, preparing the artifact path), or SYNTHESIS (producing the final artifact). State transitions are governed by the protocol. Synthesis requires explicit operator approval — implied momentum, positive feedback, and silence do not count.
Layer 2 — Minimal memory signals. GPT emits a memory signal only when omitting it would cause a real continuity loss: a confirmed decision, a real gate opened or closed, a discarded path likely to return, a source-of-truth change, or an artifact boundary. Signals are intentionally small — a type, a text, and an optional effect. GPT does not manage, rewrite, or normalize memory files.
Layer 3 — Python-owned append-only persistence. notes.py owns persistence entirely. It receives a minimal signal from GPT, timestamps it, assigns an ID and sequence number, computes a checksum, and appends the record to session_notes.jsonl. It returns a short receipt. The append log is the source of truth — GPT never reads it during live shaping, only when a recap or synthesis preparation explicitly requires it.
Final artifacts are produced only after you explicitly approve them. The artifact contains no process notes, no discarded paths, no shaping rationale, and no memory signal trace unless you specifically ask for an audit-style document.
Project files
The package contains four files. Three are part of the source you add to a ChatGPT Project; one is the Python persistence tool you run locally when you want append-only logging.| File | Role |
|---|---|
protocol.md | The behavioral contract GPT follows. Defines the three operating states, authority order, gate handling, memory signal rules, shaping principles, synthesis discipline, and failure conditions. GPT reads this at session entry. |
note-io.md | The I/O boundary between GPT and Python persistence. Defines the live write model, allowed signal types, receipt shape, readback conditions, and command set. GPT follows this contract when deciding whether and how to emit a memory signal. |
notes.py | The append-only Python persistence tool. Accepts init, append, tail, summary, and export commands. Run with python3 -S. Writes session_meta.json and session_notes.jsonl to the package directory by default. Pass --dir <path> to override the storage location. |
session_meta.json, session_notes.jsonl, and session_export.json — are produced during use. They are session-specific and should not be committed as part of the clean source package. Add them to .gitignore or keep them out of the repository entirely.
Signal types
GPT may emit a memory signal only when omitting it would cause a real continuity loss. Eight signal types are valid; no others are accepted bynotes.py:
| Type | Used when |
|---|---|
mode | An approved operating state transition occurs — for example, entering SYNTHESIS. |
shape | The candidate form of the idea moves in a new direction. |
gate | A real gate is opened, deferred, or closed. |
decision | An operator-confirmed decision is made. |
discard | A path is explicitly ruled out and likely to return. |
tension | A material conflict between active lenses is likely to recur and affect later shaping. |
source | The source of truth for the idea changes. |
artifact | An artifact boundary is established or updated. |
Gate strategies
When GPT identifies a real gate — a decision that, left unresolved, would make the final artifact wrong, unusable, or structurally misleading — it classifies the gate before acting. Four strategies are defined inprotocol.md:
| Strategy | When it applies |
|---|---|
ASK | Choosing would change operator intent, scope, authority, or a non-reversible direction. GPT asks a closed or bounded question. |
PROPOSE | One option is cheaper, simpler, safer, and reversible. GPT states the default and the reason, then continues. |
DEFER | The gate is real but not yet material to the current state. GPT marks it deferred and continues; a deferred gate is never treated as closed. |
STOP | Continuing would make the artifact wrong, unusable, or structurally misleading. GPT stops and surfaces only the blocking gate. |
Get started in minutes
Clone the repo, upload the files to a ChatGPT Project, and run your first structured shaping session.
Understand the operating model
Learn how the three operating states work together and what GPT is actually doing in each one.
Learn about real gates
Understand the gate test, how GPT classifies gates, and why fake gates are a protocol failure condition.
Python tool reference
Full reference for
notes.py — commands, flags, receipts, storage behavior, and failure handling.