what sign up
nixm nixm

← learn

the thread kept growing. the signal held.

two runs, september and november, out to turn 40. the thread's context climbed turn after turn; the signal's held between 2,000 and 3,000 tokens. where that helped, where the thread still won, and what it costs.

November 10, 2026

We compared a normal chat thread, which sends the whole transcript to the model every turn, with a nixm signal, which sends a fixed intent, a running log of decisions and the last five exchanges. Same model on both sides, scripted conversations, design registered before each run. In the first run, in September, a blind AI judge preferred the signal's replies 28 to 12. A larger second run in November, with two judges and exact-match scoring, came out level overall, 16 to 15: the signal drifted less on a single long mission (a poem 5 to 0, a 40-turn project 6 to 2), and the thread was better at exact recall (90% against 75%) and on long tangents. The signal's own context stays flat, about 2,000 to 3,000 tokens out to turn 40, while a thread's keeps growing, but today the signal costs more per turn because of a large fixed rulebook.

Why nixm works this way

nixm doesn't keep a transcript. A distilled log of the session stays with the signal for seven days, then dissolves, along with any images you uploaded. That was a privacy decision, and it has a consequence: with no thread to reread, the context has to travel with every message.

So each nixm conversation is a signal with two parts:

  • The intent: what the conversation is for, set once and fixed.
  • The running log: what has been decided so far, rewritten as the conversation moves.
A nixm signal reopened from its URL: the signal ID NIXM-DEMO-0001 in the header, and a card reading weeknight dinners for two, no dairy, one pan, spicy, friday is now a creamy coconut pasta.
A signal reopened from its link: the ID in the header, and a card written from the intent and the running log, not from a transcript. A demo signal; its card text was written for this example from the test run's log.

The model sees the intent, the log, the last five exchanges and the new message. It never sees the full history. The obvious question is whether that makes the AI worse at keeping track. We tested it instead of assuming.

The first test, in September

The design, the scenarios and the answer keys were written and committed before any run: the pre-registration.

  • Five scripted conversations of 22 to 25 user turns each: a trip plan that keeps changing, a café-name brainstorm with rejects, a long tangent followed by "where are we?", an early constraint that matters late, and planning a scene.
  • Two conditions, same model on both sides: A is a plain chat thread with the full transcript every turn; B is live nixm.
  • Three runs of each: 30 conversations and 708 turns in total.
  • Grading: check questions with answer keys written in advance, a blind pairwise judge (a separate, different AI model, shown the two replies in random order), and tokens and latency measured per turn.

The September results

MeasureThread (A)nixm signal (B)
Exact-recall checks: correct82%60%
…partly correct9%33%
…wrong9%7%
Secondary checks: correct1 of 125 of 12
Blind judge preference, final three turns1228 (5 ties)
Context tokens at turn 5 / 10 / 20 / last1,174 / 2,519 / 4,312 / 5,1211,630 / 1,913 / 1,936 / 2,307
Fixed system prompt, tokens1125,895
Mean reply time5.3 s8.4 s
September run: context sent to the model per turn. The lines cross around turn 8; after that the signal stays flat and the thread keeps growing until near the end. Measured to turn 25 only; the second run below goes to turn 40.

Where the signal won

In September, the blind judge preferred the signal's replies about two to one, and none of its stated reasons were about style. The second run, with two judges, came out level overall; that is below. The September wins clustered where conversations change under you:

  • The changing trip plan: 8 to 0 for the signal.
  • The early constraint that matters late: 8 to 1.
  • Scene planning: 7 to 0.
September run: which reply a blind AI judge preferred on the final three turns, per scenario. The signal won three scenarios, the thread won two; 28 to 12 overall, with 5 ties.

A running log of decisions keeps the current state of a plan in front of the model. A long transcript keeps every superseded version of it there too. The secondary checks point the same way: the signal was correct 5 of 12 times vs 1 of 12 for the thread.

Secondary checks answered correctly, out of 12 each: the thread 1, the signal 5.
Seeding a signal on the nixm homepage, then opening its link in a fresh tab: the intent and the decisions so far come back, and nixm answers from them. A demo signal; nixm's replies are real ones recorded in the test.

Why the goal holds

A thread keeps every version of the plan. When the user changes their mind late in a conversation, the new plan sits at the end of a transcript that still holds the old one, and the model has to work out which is current. A signal's log is rewritten as decisions change, so on each turn the model sees the plan as it stands now, next to the intent that says what the conversation is for.

nixm's reply on a phone to the final check in the dinner plan, part 1: it lists the rules (no dairy, one pan only, spicy food, rice dishes, serves two, three on thursday), then starts checking each night against them: monday has no dish yet. Part 2 of the same reply: tuesday, wednesday and thursday checked against the rules, each one fitting them. Part 3 of the same reply: friday as a creamy dairy-free coconut pasta in one pan, then monday named as the only gap.
Turn 25 of the dinner plan in the test, asked to list the rules and check each night. The rules from turns 2 and 3 are still there, and Friday is the pasta from turn 22. nixm's reply as recorded, replayed for the screenshot.

That is where the signal's clearest wins came. When a constraint set early mattered late, the judge preferred the signal's replies 8 to 1, and in seven of those eight the reason was the latest change: the user had swapped Friday's dinner for a creamy pasta, and the signal's reply built on the pasta while the thread's often still served the dish it replaced. Recall alone does not explain it: on the direct checks for that constraint, both sides scored 3 of 9. On the secondary checks, the signal was correct 5 of 12 times vs 1 of 12 for the thread; on "where are we?" alone, 1 of 3 vs 0 of 3. That scenario was not in the second run; the closest one, a running plan with a knee rule set early, went 4 to 2 to the thread.

nixm's reply when asked to make Friday a creamy pasta: a dairy-free creamy pasta with a coconut milk or cashew cream base, done in one pan, zero dairy. The phrases done in one pan and zero dairy are ringed.
Turn 22: the partner wants a creamy pasta for Friday. The reply keeps the two rules set twenty turns earlier, one pan and no dairy. nixm's reply as recorded in the test, replayed for the screenshot.

In the judge's own words: "a creamy pasta that is dairy-free, uses one pan and is spicy, so it respects all the constraints" and "It stays dairy-free by using coconut cream, cooks in one pan, and adds chili for heat."

It is not a guarantee against drift. In the long-tangent conversation the thread won 5 to 2, and 6 to 1 in the second run, so keeping the current plan in view does not, by itself, pull a derailed conversation back on course.

Where the thread won

Exact recall. When a check asked for a specific detail from earlier, the thread was right 82% of the time and the signal 60% in September, with most of the signal's misses partly right rather than wrong. The second run scored recall by exact text match, with no AI grading it: 90% for the thread, 75% for the signal.

In September, the café brainstorm exposed why. The thread got its recall checks 9 of 9; the signal got 0 of 9. On the very first message, nixm wrote its own opinions into the log as if the user had decided them, noting names as cut that the user had never rejected. Once that exchange scrolled out of the last five, the log was the only record, and it was wrong: names the user still wanted were gone. The thread also won that scenario on the judge's preference, 6 to 3, and won the long tangent 5 to 2.

What we changed

Two earlier attempts at a fix did not hold. The one that worked is simpler. The log now holds only what the user said, including every item they list, and nixm's own opinions stay in its reply instead of being written down as decisions.

In a targeted rerun of the brainstorm, the first message kept all ten names in 5 of 5 tries, and every name the user cut was recorded correctly through turn 15 in both 15-turn runs. The second run caught it slipping once more, on a newer version of nixm, and one of the November fixes below closed it: all ten names kept on the first message in 7 of 7 live tries.

One known limit remains: when a user cuts a name, the log drops it rather than recording why it was cut.

The second run, in November

In November we ran it again, built to be harder on the signal and to rely less on AI grading. The design was registered before any run, like the first.

  • Five scenarios, two runs each: the café brainstorm and the long tangent from September, plus a birthday poem revised over eight turns, a 40-turn website project whose budget changes three times, and a running plan where the user comes back from a detour with only "ok so what now?". 30 conversations, 678 user turns.
  • A third side: a plain thread that also carries nixm's rulebook, so the rulebook and the memory can be told apart.
  • Recall scored without AI: planted facts and stale values were checked by exact text match.
  • Two judges for intent: AI models from two different companies, each pair judged twice with the order swapped. A win counts only when both orders agree, and the headline counts only items where both judges agree.
Second runThreadnixm signalThread + rulebook
Planted facts found (exact match)90%75%84%
Old values presented as current, of 48 checks050
Replies both judges rated as serving the goal, of 42201925
Head to head with the signal, both judges15–16, 3 ties—14–12, 3 ties
Context tokens at turn 204,8722,370 est.8,041
Context tokens at turn 409,1032,798 est.18,071
Fixed system prompt, tokens11about 23,400about 23,400
Mean reply time5.3 s8.8 s8.2 s
Context sent to the model per turn, in the second run's 40-turn conversation. From turn 8 the signal sends less than either thread, and it stays between about 2,000 and 3,000 tokens to the end while both threads keep growing.

Overall the judges called it level: the signal 16, the thread 15, with 3 ties. September's two to one did not repeat. Where it split is the useful part.

Which reply both judges preferred, per scenario, in the second run. The poem and the 40-turn project went to the signal; the brainstorm, the ambiguous return and the long tangent went to the thread. 16 to 15 overall, with 3 ties.

On a mission, the signal drifted less. The birthday poem went 5 to 0 for the signal and the 40-turn project 6 to 2. Both build one thing across many changes: an age corrected, a budget raised and cut and raised again, a launch date moved, a colour swapped, a page added and then dropped. In the 40-turn project no side presented an old value as current.

nixm's finished birthday poem, titled for odile, turning 88: peonies in the first line, Brittany, the Atlantic, eighty-eight years, six lines in all.
Turn 8 of the poem. Every change the user made is in it: the age corrected from 87 to 88, peonies, Brittany, a flower in the first line, under 12 lines. nixm's reply as recorded in the test.
nixm's reply after the user says sorry, where were we: the dog grooming site's pages, prices from 45 pounds for small dogs, teal and mustard, launch May 16, a 20% first-groom offer, budget 900 pounds.
Turn 30 of the 40-turn project, right after a detour about the user's dog and the vet: "sorry where were we". The launch is May 16 (moved from May 2), the colours teal and mustard (coral was dropped), the budget £900 (it rises to £1,500 next turn), and £45 was said once, at turn 6. nixm's reply as recorded in the test.

The thread was better at exact recall and on tangents. It found 90% of the planted facts against the signal's 75%, and won the long tangent 6 to 1, the ambiguous return 4 to 2 and the brainstorm 3 to 2. The weak spot we named in advance showed up too: a price said once, at turn 6, and asked for at turn 38, came back right in one run and wrong in the other.

nixm's reply: cheapest price is 45 pounds, for small dogs, and the business is sudsy paws. nixm's reply in the other run: cheapest price is 15 pounds, that's the nail trim, and the business is sudsy paws.
The same question at turn 38, "what did I say my cheapest price was?", in the two runs. The first reply is right: £45, said once at turn 6. The second gives £15, a price from the prices page nixm itself wrote at turn 34. nixm's replies as recorded in the test.

The rulebook does part of the work. The thread that carried nixm's rulebook had the most replies the judges rated as serving the goal, 25 of 42, beat the signal 14 to 12 head to head, and tied the plain thread. Some of what looked like the signal's edge in September comes from the rulebook, not from the log.

Four memory bugs, found and fixed. The second run surfaced four ways the signal's log lost track:

  • Limits: a limit the user set, "under 250 words, warm", was followed once but never written into the log, so a few turns later nixm said no limit had been given.
  • Drafts: a draft written before a long tangent fell out of the last five exchanges, and nixm said no draft existed.
  • Killed names: a name the user had cut came back into the log a few turns later, copied from an older list.
  • First-message lists: on the first message, ten names were cut to an invented shortlist of five.

Each was traced to its cause in nixm's instructions or in what it sends back to the model, fixed, and tested again live on fresh signals: the limits held in 3 of 3 runs, the draft was rebuilt instead of denied in 2 of 2, all ten names were kept on the first message in 7 of 7, and no cut name came back in 2 of 2. The second-run numbers above were measured before these fixes.

nixm's reply to where are we: a cover letter for a junior UX designer role at Maple Health, three years in retail management, a UX bootcamp finished in June, a food bank intake redesign from 12 minutes to 5. No word limit or tone.
Before the fix: turn 18 of the cover letter, after a coffee tangent, "where are we?". The role, the company, the background and the 12-to-5-minute result are there; the limit set at turn 4, under 250 words and warm, is not. Since the fix, the log keeps it. nixm's reply as recorded in the test.

A signal is for one mission. Its intent is locked from the moment it's set, and that is what kept it on course across forty turns of changes. A long, wandering conversation is where a plain thread does better. When the mission is done, or the conversation turns into a different one, start a new signal: on the nixm homepage, say what the new one is for, or ask for a link, and nixm seeds it.

What it costs

Plainly: the signal is more expensive per turn today. It carries a fixed rulebook, about 26,000 tokens in September and about 23,400 now, which the plain thread does not, and its replies were slower: 8.4 seconds on average against 5.3 in September, 8.8 against 5.3 in November.

The part that stays small is the conversation's own context. In September it sat around 2,000 tokens from turn 10 to turn 25, against a thread that peaked at about 6,000. In November we went to turn 40: the signal stayed between about 2,000 and 3,000 tokens, while the plain thread climbed to about 9,100 and the thread carrying the same rulebook to about 18,000. So the signal's extra cost is the fixed rulebook, not the conversation: its cost per turn barely moves as the work goes on, while a thread's rises every turn. Against a thread that carries the same rulebook, the signal sends less from turn 8 on. We measured to turn 40, so we are not claiming where the curves go after that.

The limits of this test

  • Five scenarios per run, three runs each in September and two in November, is enough to see a pattern, not to settle a question.
  • The judges are AI models, not people: one in September, two from different companies in November.
  • Both sides used one model; another model might behave differently.
  • Conversations ran to 25 turns in September and 40 in November.
  • The November numbers were measured before the four memory fixes; the fixes were tested on their own, not with a full rerun.

The September pre-registration, with its scripts and answer keys, is linked above with the date it was first committed, so the design can be checked against the results.

Why this matters beyond nixm

Most AI chat keeps everything and rereads it every turn, because that is the easy way to keep context. This test suggests you do not need the whole transcript to keep a conversation on track: on a single mission, a decision log drifted less. It is weaker at word-for-word recall and on long tangents, and it has a cost you can see. For a tool whose point is that there is no transcript to keep, that is the trade we would rather make in the open. How nixm's storage works is in what ephemeral AI chat is, and how it compares with other assistants is in what each AI chatbot keeps. If you are here because a long chat hit its limit, what ChatGPT's maximum length message means covers what to do next.

ask anything. nothing is kept.

try it now

ask anything. nothing is kept.

no account, no transcript kept. even the signal dissolves in seven days.

seed a signal → free · opens in this tab

frequently asked

Does an AI chat need the whole transcript to keep track?

Not in this test. In the first run, a fixed intent plus a running log of decisions was preferred by a blind AI judge 28 to 12, especially when plans changed. A second, stricter run with two judges came out level overall, 16 to 15: the signal did better on a single long mission, the full transcript on exact recall and long tangents.

What is a nixm signal?

A conversation's fixed intent plus a running log of what has been decided. The model sees those, the last five exchanges and the new message, never the full history. The signal dissolves seven days after it is created.

Where did the full transcript do better?

Exact recall: 82% correct against 60% in the first run, and 90% against 75% in the second, scored by exact text match. It also did better on long tangents. In September the signal's log recorded nixm's own opinions as if the user had decided them; that and four memory bugs found in November are fixed.

Is the signal approach cheaper?

Not today. Its own context stays around 2,000 to 3,000 tokens out to turn 40, but it carries a fixed rulebook of about 23,400 tokens, so each turn costs more and replies were slower: 8.8 seconds against 5.3 in the second run.

Was the test fair?

The design and answer keys were committed before each run, both sides used the same model, and the judges saw replies blind. The second run scored recall by exact text match and used two judges from different companies, each pair judged in both orders. It is still a small test, and the judges are AI models, not people.

Does nixm keep a transcript?

No. A distilled log of the session stays with the signal for seven days, then dissolves, along with any images you uploaded.

Can I try nixm without an account?

Yes. Open nixm.ai and seed a signal, with no sign-up and no transcript kept.

Enter your email and receive a link to sign in. No passwords, just a secure signal to connect.

why connect

  • 20 free credits to start
  • your signals, remembered
  • buy more credits anytime
  • image & video generation features
or