Qontext

A quipu-inspired conversational memory for language models. Pure Python, one file, no dependencies, Python 3.8+.

The Inca recorded information as knots on cords — not a transcript of everything said, but a compact encoding of what was worth keeping. Qontext stores a conversation the same way: short, self-contained facts ("knots"), and pack(query, budget) returns only the ones a given question needs.

pip install qontext-memory

from qontext_memory import QontextMemory

memory = QontextMemory()
memory.observe("user", "People call me Marta and I work as a nurse.")
memory.observe("user", "The report is due March 3rd, hard deadline.")

memory.pack("when is the report due?", 300)
# 'the report is due March 3rd'

What it is for

A long conversation does not fit in a small context window, and sending all of it is expensive even when it does. Qontext keeps a fixed-size summary that does not grow with the conversation.

Measured on 800-turn conversations whose filler is real human dialogue, against a 12B model, three seeds:

Three seeds cannot separate 9.3 from 9.7 — that is a tie. The 119x reduction is the result: the same answers for one percent of the prompt.

Re-verified under speakers="all" (the current default; the table above was originally measured under speakers="user", the previous default). Same command, same three seeds, same 800-turn conversation:

The pack side — the one speakers= actually changes — held almost exactly: same 9.7 accuracy, same single miss (still the assistant's name question, still answered "Terry Graham"), 117 tokens against a previously-measured 116. The full-transcript side improved from 9.3 to 10.0, but that arm never touches QontextMemory at all, so this isn't something the default change gets credit for — most likely run-to-run model variance (sampling is not deterministic at temperature 0.2), not re-investigated further. The 119x figure and the tied-accuracy claim both stand under the current default.

On smaller models the picture is stronger and more conditional. A 9B scores 4.3/10 from a transcript full of plausible wrong answers against the pack's 8.0; a 4B scores 4/10 against 10/10. The honest statement is that context reduction improves accuracy when the context holds plausible wrong answers and the model is at or past its ability to choose among them — distractor density sets the difficulty, model capability sets the threshold. On a capable model you get the saving, not the accuracy.

Qontext_Study_Report.pdf is the full write-up, corrections included.

Design rules

Three rules, each paid for with a failed run:

1. A knot must name its subject. "I work as a nurse" is unanswerable once cut from the conversation; "the user works as a nurse" survives alone.
2. The payload is the point. "People call me" without "Marta" is a knot tied around nothing. Never trim the name, day, place or number out of an entry.
3. Fewer, better knots. Density comes from selection, not truncation. Cutting entries short to meet a budget destroys the meaning the budget exists to protect.

What is here

python -m unittest -v test_qontext_memory
python live_agent.py

Roleplay

qontext_rp.RPMemory is the same idea under roleplay's assumptions. The chat extractor admits a sentence on a marker or a payload — a number, a date, a capital — and roleplay's most important sentences have none of those. "I am at the harbour now, not the shrine" was dropped entirely, leaving only "Plans changed". The RP build treats state as payload: place, possession, condition, kinship and commitment, recognised structurally because the nouns of a setting cannot be known in advance.

Measured on 11 public roleplay logs, turn-shaped queries at budget 1200: 4.1% of needed facts carried by the chat build, 9.9% by the RP build. That is a doubling, and it is still one fact in ten — the turn-shaped task is mostly unsolved, and now measured rather than assumed.

SillyTavern extension

Copy extension/qontext/ into SillyTavern/public/scripts/extensions/third-party/. Three modes: regular (off), augment, replace.

The vocabulary tables are generated from the Python source rather than retyped, and extension/parity.mjs replays scenarios through both implementations and fails on any divergence. It found a real bug during the port.

The word weave, and why it is off

qontext_weave.py links words by co-occurrence so a query can reach a knot it shares no word with. The mechanism works — seeded from WikiText-103 it knows vampire → slayer, dracula and allergic → anaphylaxis, asthma. Measured on real logs it was neutral to slightly harmful at every setting. Kept, wired, documented, and None by default.

Honest limitations

• The ten benchmark questions are ours, even where the conversation is not.
• Measurements come from one machine and two local quantised models.
• Answers are scored by keyword match, not judgement.
• The roleplay result has no model-in-the-loop validation, only a proxy.

Licence

PolyForm Noncommercial License 1.0.0 — see LICENSE.

Free to use, modify, and share for research, teaching, personal projects, and by non-profit, educational and government organisations. Commercial use is not permitted. For a commercial licence, open an issue.

This is source-available rather than open source in the OSI sense: the "no commercial use" restriction is what the OSI definition excludes. Calling it source-available is the accurate description.

The DailyDialog filler used by bench/long_bench.py --filler daily is not redistributed here — DailyDialog carries its own licence, and bench/build_filler.py rebuilds the file from the original dataset in one command.