Updated 21 October 2026. Since May, xAI has published much more of the For You code, including the full filter list and the configuration defaults. The biggest correction: we said posts older than 48 hours could keep surfacing out of network. They cannot, because AgeFilter removes every candidate older than 48 hours, whatever source found it. We also fixed file paths, replaced guessed per-author caps with the real ones and added the age signals the newer code shows.
“Recency boost” is one of the most repeated pieces of Twitter algorithm advice. The open-source For You code tells a more precise story about Twitter algorithm recency. There is no hand-set recency multiplier in the score. There is a hard window: a post older than 48 hours is removed from the For You candidate pool before scoring, whichever system found it. Inside that window, age still matters in three softer ways, and this article walks through each with file paths you can check.
If you came here asking how long tweets stay in the feed, the short answer for For You is 48 hours. The longer answer explains why most of a post’s reach arrives early, and what that means for when and how often you post. It pairs with the series hub on the 2026 Twitter algorithm and how often to post on Twitter.
The short answer: 48 hours, for every source
The README’s table of pre-scoring filters lists AgeFilter with a plain description: “Posts older than 48 hours”. The code matches. AgeFilter in home-mixer/filters/age_filter.rs keeps a candidate only if its age is at or below a maximum, and drops it if the age cannot be read. The maximum is MAX_POST_AGE = 48 * 60 * 60 seconds in home-mixer/params/config.rs. The filter is wired into the post pipeline in home-mixer/candidate_pipeline/phoenix_candidate_pipeline.rs, fourth in the list of pre-scoring filters, after duplicates are merged and before anything is scored.

The README’s pre-scoring filter table, including AgeFilter: “Posts older than 48 hours”. Source: github.com/xai-org/x-algorithm README, captured 21 October 2026.
Placement is the important part. Filters run on the combined pool, after every source has returned its candidates. So the 48-hour cut applies to posts from accounts the viewer follows and to posts found out of network alike. A post is either inside the window or it is not in For You at all.
What we got wrong in May: no out-of-network afterlife
Our May version said that after Thunder’s 48-hour window closed, a post could keep surfacing through out-of-network retrieval for days or weeks, until a longer tweet-mixer cutoff. That is not what today’s code does. Every source either has its own limit of 48 hours or less, or is caught by AgeFilter anyway.
Age limits by For You candidate source (21 October 2026)
Source | Its own age limit | Where |
|---|---|---|
Thunder (accounts you follow) | Keeps posts 172,800 seconds (48 hours) by default |
|
SimClusters (out of network) |
|
|
Popular posts (shared list) | 24-hour window, ranked by views with an 8-hour half-life |
|
Tweet mixer |
|
|
Phoenix retrieval and cached posts | No separate cut in home-mixer; AgeFilter still applies |
|
Two notes. The Thunder retention value is a command-line default, so the production setting is not in the repo, but AgeFilter enforces 48 hours either way. And the Phoenix retrieval index in phoenix-rankall/ keeps some longer windows, for example for video, yet anything it hands to For You still passes through AgeFilter.
So how does an older post ever come back? The code shows a few narrow routes. None of them ranks an old post as itself:
A fresh repost or quote. AgeFilter reads the age from the candidate’s own post ID, and for a repost Thunder uses the repost’s own ID (
home-mixer/sources/thunder_source.rs). Our read: a repost starts a new 48-hour clock for the reposter’s followers. The seen filters also check the original post’s ID, so people who already saw the original do not get it again.Push-to-home. When a request names a specific post to push to Home, that post is fetched directly and pinned at the top (
home-mixer/sources/push_to_home_source.rs). It skips the post pipeline, so AgeFilter does not apply. Our read: this is notification territory, not ranking.Other surfaces. Profiles, search and the Following tab run separately. The Following feed’s night-owl source, for instance, looks back up to 365 days (
RECALL_MAX_AGE_SECSinhome-mixer/sources/following_night_owl_source.rs). This article is about For You only.
Is there a recency boost? No multiplier, but age still matters
The ranking score has no age term. The README describes it as Final Score = Σ (weight_i × P(action_i)), and compute_weighted_score in xai-value-model/scoring.rs sums weighted predictions with no recency factor. There is no RECENCY_WEIGHT anywhere in the repo. Those weights multiply predicted probabilities, not counts, and the repo’s own comments warn that reading them as “one report cancels 468 likes” is wrong.
Our read on why the recency-boost idea survives: most posts really do get most of their reach early, and that looks like a boost from the outside. The code explains the pattern without one. A hard 48-hour cut ends every post’s run, the in-network store hands over its newest posts first, and the seen filters stop a follower being shown the same post twice. Put those together and early hours matter a great deal, even though nothing in the score rewards freshness by name.
Age still shapes what you see, in three ways the code makes visible.
1. Phoenix can learn from post age
Our May version said the model had no “this post is fresh” input. The Phoenix training code published in August has one. compute_post_age_bucket in phoenix/xrex/models/recsys_model.py turns the gap between impression time and post creation into hourly buckets up to 4,800 minutes (80 hours), and the training config in phoenix/xrex/configs/xrecsys.py turns the post-age feature on. In plain English, the model can learn how age changes the chance of a like, a reply or a long read. That effect is learned from data rather than hand-set, and the code does not tell us how big it is.

Phoenix’s post-age feature: hourly buckets up to POST_AGE_MAX_MINUTES = 4800. Source: phoenix/xrex/models/recsys_model.py, captured 21 October 2026.
2. Thunder serves newest first
Thunder is the in-memory store of recent posts from accounts you follow. It sorts the posts it finds by creation time and returns the newest first (score_recent in thunder/thunder_service.rs). Home-mixer asks for up to 1,200 (ThunderMaxResults in home-mixer/params/param.rs). For a viewer who follows many busy accounts, an older post inside the 48 hours can be pushed out of the in-network pool by newer ones before ranking starts.
3. The new-author boost lasts two hours
The new-author boost in home-mixer/scorers/author_cold_start.rs lifts at most one post per feed request to about the 16th slot. It only considers original posts that are 7,200 seconds old or less (ColdStartMaxPostAgeSecs), with fewer than 200 Home views (ColdStartImpressionThreshold), from authors with 50,000 followers or fewer (ColdStartFollowerCap). For smaller accounts, the first two hours are a separate, shorter window inside the 48.
Thunder: how the in-network window works
thunder/posts/post_store.rs holds recent posts in memory as they are published, in three lists per author: original posts, replies and reposts, and video posts. A trim job removes posts older than the retention setting, 48 hours by default. Our May version cited a top-level thunder/post_store.rs; the file now lives under thunder/posts/.
When home-mixer asks for a viewer’s in-network posts, Thunder walks the viewer’s follow list and takes a capped number from each author. It reads at most 10,000 accounts from that list (MAX_INPUT_LIST_SIZE), so very large follow lists are cut short. The per-author caps are in thunder/config.rs:
MAX_ORIGINAL_POSTS_PER_AUTHOR = 50MAX_REPLY_POSTS_PER_AUTHOR = 30(replies and reposts)MAX_VIDEO_POSTS_PER_AUTHOR = 100(for video requests)
In May we said these values were not public and described them as limits on each viewer’s personal store. Both points were wrong. The caps are public, they apply per author on each request, and they are generous. Our read: 50 original posts inside 48 hours is far beyond normal posting, so for almost every creator the per-author cap never binds. The limits that do bite are the newest-1,200 cut above and the author diversity discount further down the pipeline.
The seen and served filters
Once a post has been shown, For You works hard not to show it again. Three pre-scoring filters handle this, and the README now documents all three. In May we marked the third as inferred.
Repeat-prevention filters in home-mixer/filters/
Filter | What it checks |
|---|---|
| Post IDs the app reports as seen, plus impression bloom filters for the Home timeline |
| A second record of impressions |
| Posts served recently; the defaults are the last 10 minutes and up to 100 IDs ( |
Each filter checks the post itself, the original behind a repost and the post a reply answers (related_post_ids_iter in home-mixer/util/candidates_util.rs). Thunder is also handed the seen list and leaves those posts out at the source. Our read: in For You, each follower effectively gets one showing of each post, and a repost of something they already saw does not count as new.
A post’s life in For You, hour by hour
Put together, here is what the code does to one original post over time. The hours come from the code; the right-hand column is our read.
One original post in For You, by age
Age | What the code does | Our read |
|---|---|---|
0 to 2 hours | Thunder holds it for followers; Phoenix’s age feature sees it as new; if the author has 50,000 followers or fewer and the post has under 200 Home views, it can be the one post lifted to about slot 16 | The busiest window for smaller accounts |
2 to 24 hours | New-author boost over; it competes on predicted actions; seen filters remove it for each follower who has been shown it; if the author is among the most-followed accounts, it can join the shared popular-posts list | Most followers who will see it in For You probably see it here |
24 to 48 hours | Popular-posts window closed; still eligible from Thunder and out-of-network retrieval, but newer posts crowd the newest-1,200 in-network cut for heavy-follow viewers | A thin tail |
After 48 hours | AgeFilter removes it from For You for everyone; Thunder trims it | Only a fresh repost, a quote or push-to-home brings it back |
Our May version described a slow handover from in-network to out-of-network reach after 24 hours, and a long tail beyond 48 hours. Neither shows up in the code. Both lanes close at the same 48-hour line.
What this means for cadence
Recency and frequency meet in one place. Every post gets its own 48-hour window, but posts from the same author compete with each other inside it. Per-author caps in Thunder are not what limits you; the author diversity discount is. In xai-value-model/scoring.rs, each extra post from the same author in one feed request is multiplied by (1 - floor) * decay^n + floor. The code defaults in vm-ranker/params.rs are decay 0.5 and floor 0.25, but they are not in the synced production file, so the production values are unclear from the code. We walk through the maths in how often to post on Twitter.
Our read: a short window rewards timing more than volume. Posting when your audience is online (try our best time to post tool) puts a post into followers’ pools while it is fresh, and spacing posts out gives each one a better chance of landing in a different feed request from your others.
What to do with this
Four moves follow from the code. All four are our read, not rules the code states.
Treat 48 hours as the whole For You life of a post. If an idea deserves a second run, write a new post with a new angle rather than waiting for the old one to resurface.
Be around for the first two hours. For smaller accounts that is the new-author boost window, and the replies you earn early count while the post is still in everyone’s pool.
Earn reposts and quotes. They are the main way an older idea gets a fresh 48-hour clock in front of new people.
Do not chase volume to beat the window. More posts inside the same 48 hours mostly compete with each other through the author diversity discount.
VoiceMoat’s point of view is that consistent, recognisable content from a real person tends to earn the positive actions the model predicts (replies, dwell, shares, follows) and avoid the negative ones (not interested, mute, block, report). That is our read, not something the code says. What the code does say is simpler: inside 48 hours, each post gets about one showing per follower, and the post itself has to earn the rest. For the full ranking path, start with the series hub; for what a post competes with once it reaches the page, see how ads, Who to Follow and prompts fill the feed; for reach beyond your followers, see out-of-network retrieval.
How long do tweets stay in the Twitter For You feed?
48 hours. AgeFilter in the open-source home-mixer code removes any candidate older than 48 hours before scoring, whichever source found it (MAX_POST_AGE in home-mixer/params/config.rs). After that, a post only returns through a fresh repost or quote, a push-to-home request, or other surfaces such as profiles and search.
Does the Twitter algorithm give a recency boost?
Not as a multiplier. The ranking score is a weighted sum of predicted probabilities with no age term. Age still matters in three ways: the Phoenix training code includes a post-age feature, Thunder returns followers' posts newest first, and the new-author boost only covers posts up to two hours old.
Can an old tweet come back in the Twitter feed?
Not as itself after 48 hours. A repost carries its own post ID, so it gets a new 48-hour clock for the reposter's followers, though people who already saw the original are filtered out. Push-to-home can pin a specific post at the top, and the Following tab looks back further.
What is Thunder in the Twitter algorithm?
Thunder is the in-memory store of recent posts from accounts you follow. It keeps posts for 48 hours by default, returns up to 50 original posts and 30 replies or reposts per followed author, and sorts by newest first. Home-mixer asks it for up to 1,200 posts per request.
Does posting time matter on Twitter?
Our read is yes, because the window is short. A post has 48 hours in For You, the new-author boost covers only its first two hours, and Thunder favours the newest posts. Posting when your audience is online gives a post more of its window in front of them.
This article was drafted with AI assistance and human-edited by Prateek Singh. All technical claims are sourced to the xai-org/x-algorithm repository; file paths are cited inline.
More in the series

Twitter Algorithm
The Twitter algorithm in 2026: how the open-source ranker works and why your voice matters
Twitter's For You feed runs on open-source code in the xai-org/x-algorithm repo on GitHub, first published in January 2026 and updated almost every weekday since 13 August. A transformer called Phoenix predicts 25 actions for each post, published weights turn them into one score, and author diversity, an out-of-network discount, a new-author boost and a DPP reranker follow. This hub walks the current code in plain English and explains why consistent, recognisable content still matters for creators.
14 min read

Twitter Algorithm
How the Twitter Phoenix algorithm sees creators: hashed authors, semantic IDs and viewer history
Phoenix, the model behind the Twitter For You feed, never reads your "voice". It reads hashed author IDs, semantic IDs that describe each post, and each viewer's last thousand or so actions. This refresh walks the production code published since August 2026 and labels clearly where our creator advice is interpretation rather than code fact.
12 min read

Twitter Algorithm
Twitter out of network: how your posts reach people who don't follow you
Out-of-network posts reach strangers on the Twitter For You feed through three live lanes: Phoenix retrieval, SimClusters and a shared popular-posts list. This refresh rebuilds the article from the October 2026 code, including the 0.75 out-of-network discount, the filters that keep replies and reposts from travelling, the extra visibility rules for recommendations and the new-author boost for small accounts.
11 min read

Twitter Algorithm
How often to post on Twitter: author diversity, DPP and your third post of the day
How often should you post on Twitter? The open-source For You code answers part of it: each extra post from the same author in one feed request is multiplied by (1 - floor) × decay^n + floor, and a determinantal point process then picks a varied set from the top-scored posts. The repo's code defaults are decay 0.5 and floor 0.25, but production values are not confirmed. This refresh fixes our May file paths and explains what the maths does and does not say about posting frequency.
10 min read
The whole series
The Twitter/X algorithm in May 2026: all ten articles