TinyMemes

Makes an agent's final reply more fun when the chat can take it. The reply comes back in the user's own language and register, with regional slang and, sometimes, a reaction meme. A serious or formal chat is never touched.

Every reply goes through the same three steps. Two learners run afterwards, off the reply's path:

conversation + reply
      │
      ▼
 ┌─────────┐  ONE Jev call: intents, frankness, playfulness, seriousness,
 │  Read   │  user's language, does the reply already match the user,
 │         │  which slang fits, which meme fits (with Jev's probability)
 └────┬────┘
      ▼
 ┌─────────┐  pure policy → 0–10 rating → Off / Light / Spicy / Unhinged
 │  Rate   │  serious → 0 · bad news ≤ Light · meme cooldown · weak meme pick → none
 └────┬────┘
      ▼ (not Off)
 ┌─────────┐  reply already matches the user → keep wording, attach meme (no LLM)
 │  Remix  │  otherwise → ONE rewrite call (language rule, slang menu, meme slot)
 └────┬────┘           → keep code/links/length, repair duplicated lines
      ▼
   reply out ─────────►  background: slang research · meme research + vetting

A remixed reply costs one Jev call plus at most one LLM call. In live tests, a meme-only remix took about 0.6–0.9 s and a full rewrite about 4–5 s. A reply rated Off costs only the Jev call (about 0.6 s).

1. Read: one Jev call

Jev (tinyinference-decisions) reads the conversation as the user saw it. Earlier remixed replies are tagged as the bot's own style, and memes in them appear as [meme: title]. Jev answers typed questions about the conversation and the reply. It never writes text.

Question	Type	Used for
chat_intent, reply_intent	Choice (celebration, frustration, confusion, banter, grind, bad news, curiosity, gratitude)	rating adjustments, meme moment
frankness	Score (formal … unfiltered)	rating
playful	Noul	rating
serious	Noul (grief, health, layoffs, …)	the safety gate
user_language	Choice (english, hinglish, devanagari, other)	language rule and slang menu
reply_matches_user	Noul	meme-only mode
slang_best / slang_enough	Choice / Noul over the slang index	rewrite hint, slang research trigger
meme_best	Choice over the meme catalog + none_fit	which meme, if any; meme research trigger

Jev's frankness and playfulness questions are about the user's tone, so the bot's own slang in earlier replies doesn't push the rating up.

2. Rate

score = frankness × 6 + playfulness × 4, then +1 for banter or celebration, or −0.5 for grind or confusion.

Tier	Rating	Voice	Memes
Off	0–2	untouched	0
Light	3–4	a few casual words	0
Spicy	5–7	slangy	0
Unhinged	8–10	full group-chat	1

• Serious gate: serious ≥ 0.35 sets the score to 0, however jokey the user is.
• Bad news is capped at Light.
• Meme cooldown: no meme if one went out in the last 3 replies.
• Weak pick: if Jev's probability for its meme pick (meme_p) is under 0.5, it counts as none_fit. No poorly matched meme goes out.

3. Remix

Meme-only mode

When Jev's reply_matches_user is at least 0.6, the agent's reply already sounds like the user. It isn't rewritten: Jev's meme is attached after the first paragraph, and no LLM call is made. This is the common case on playful chats with a model that already mirrors tone.

Rewrite

One LLM call. The rewrite sees only:

• the reply;
• the user's latest message, for language and script;
• the language rule from user_language: English stays English, Hinglish stays Hinglish, Devanagari stays Devanagari, any other language is mirrored;
• a slang menu that follows the language: the region's slang index for Hinglish and Devanagari, English internet slang for English, none otherwise. Each term comes with its meaning;
• Jev's best-fit term, and the slang used in recent replies (so it varies);
• at most one meme candidate (Jev's pick, with its meaning), placed as [[meme:1]].

It never sees earlier messages, so it can't answer or blend them in.

After the call:

• Protected content: a rewrite that changes a fenced block, inline code or a URL, or more than doubles the length, is discarded and the original sent.
• Duplicate repair: a substantial line the rewrite repeats loses its extra copies. The copy mirroring the original line is kept.
• Meme markers resolve to ![title](url); unknown or extra ones are dropped. A meme is never forced in.

Gaali in a rewrite is not grounds for rejection. The prompt asks for none.

4. Learning in the background

Neither learner runs on the reply's path. Both persist as JSON snapshots.

Slang research (SlangIndex)

• Trigger: Jev's slang_best is none_fit, or slang_enough < 0.5, on a remixed reply. Skipped for English and other-language writers.
• Query: about the reply itself ("Hinglish slang that would fit naturally in a casual reply to this message: …").
• Vetting: the source must be a page the search cited; the domain must not be adult or spam; the term must be a word or short phrase (not a sentence), at most 3 per page, and clear the blocklist. Then one batched Jev check (p ≥ 0.6) that it's genuine, inoffensive regional slang with that meaning.
• Ranking: intent fit, corroboration across searches, and usage in replies. The top 18 for the tier are offered.
• Limits: the same query isn't repeated within 7 days; 50 searches per region per day.

Meme research and vetting (MemeIndex)

• Trigger: a meme was allowed (Unhinged, cooldown clear) but Jev found nothing in the catalog that fits (none_fit, or a weak pick).

• What to search for: the moment the user is in. The chat model turns the user's message into a short, generic meme concept: a known catchphrase or trope like "Sharma ji ka beta" or "emotional damage", never names, places, numbers or other personal details. The search is "<concept> Indian meme GIF site:giphy.com".

• Reading GIF pages (GiphyPageResearcher): search engines return pages, not media, so each GIPHY GIF page is read for what GIPHY itself publishes there:

  • the media link, switched to the small 200.webp rendition, from allowed hosts only, and checked to be a real image under 1.5 MB;
  • GIPHY's own content rating;
  • the GIF's tags.

  No GIPHY API key is needed.

• Vetting, in order:

  1. Rating: GIPHY's rating must be G or PG. This is the image-level check: GIPHY moderates the media itself.
  2. Topic filter: slug, title and tags must not touch politics, religion, sex, violence or gaali. In a live test this caught a "Shocked India" GIF whose tags were modi, vote chor, cjp protest.
  3. Meaning: the chat model names each GIF and says when people post it, tied to the moment it was found for.
  4. Jev, p ≥ 0.75: a widely recognised meme from the region's internet culture with that meaning, and not political, religious, sexual, violent, mocking a community, or satire of a real politician, journalist or public figure.

• Use: there's no approval step; the filters are the gate. GIFs that pass join the catalog Jev picks from, minus memes sent in the last 6 replies. Usage is tracked.

• Limits: the same query isn't repeated within 3 days; 20 searches per region per day; concurrent steps for the same query collapse into one.

Regions

Region::india() is the default. It sets flavour, not language:

• Curated slang: a Hinglish list with meanings, which seeds the slang index.
• Meme catalog: 13 hand-checked Imgflip templates, each with its meaning and intents (Hera Pheri, Sacred Games, Panchayat, Taarak Mehta, Moye Moye…).
• Search phrasing for slang and meme research.
• Blocklists: a gaali list, plus topics a learned GIF must not touch.

Region::global() is the plain English pack. Packs are plain data, so a host can extend them.

Guarantees

• Fails open. MemeEngine::process never errors. On any failure the original reply comes back with Outcome::skipped set.
• Serious chats stay serious. The serious question gates everything.
• Code and links survive. See Rewrite.
• No unvetted media. Only the curated catalog, or learned GIFs that passed every check.

Use

Every backend is a trait, so a host plugs in its own stack: Evaluator (Jev), ChatModel, MemeSource, SlangResearcher, WebSearch, MemeResearcher.

let engine = MemeEngine::builder(jev, chat_model)
    .slang_index(slang_index)
    .meme_index(meme_index)
    .researcher(Arc::new(SearchResearcher::new(search.clone(), chat_model.clone())))
    .meme_researcher(Arc::new(GiphyPageResearcher::new(search, http)))
    .learn_inline(None)
    .build();

let out = engine.process(&conversation, &reply).await;   // never fails
if out.wants_more_memes() {
    // in the background
    engine.learn_memes(reading.reply_intent, &user_message).await?;
}

Outcome carries the reading, the rating, the remix (mode, memes, slang_used, duplicates_removed) and what was learned.

No Jev? The LLM answers Jev's questions

LlmEvaluator wraps any ChatModel and answers the same Choice, Score and Noul questions as JSON, validated like a Jev response. The whole engine (reading, slang and meme vetting) works with no Jev route.

Configuration: host first, TINYMEMES_* env overrides

Variable	Meaning
TINYMEMES_OPENROUTER_KEY	OpenRouter for the rewrite model and slang research (reasoning disabled)
TINYMEMES_MODEL	rewrite model (default deepseek/deepseek-v4-flash)
TINYMEMES_JEV	auto, typesafe, openrouter, openjev, or llm
TINYMEMES_TYPESAFE_KEY / TYPESAFE_API_KEY	TypeSafe System One
TINYMEMES_OPENJEV_KEY / OPENJEV_API_KEY	OpenJEV

Inside a host, anything unset falls through to the host. Standalone, MemeEngine::from_env() builds everything from these variables (and accepts OPENROUTER_API_KEY):

OPENROUTER_API_KEY=… cargo run --example remix -- chat.json
TINYMEMES_JEV=llm OPENROUTER_API_KEY=… cargo run --example remix

Policy knobs live in RatingPolicy (tiers, serious cutoff, cooldown, meme_only_above, meme_min_p), IndexPolicy (slang) and MemeIndexPolicy (memes).

In OpenHuman

OpenHuman vendors this repo at vendor/tinymemes, with a host adapter in crates/openhuman-core/src/tinymemes/. It supplies:

• Inference: OpenHuman's provider for the summarization role (managed or BYOK), with reasoning off for tinymemes' calls only.
• Jev: the OpenHuman-managed route; OpenHuman's LLM answers its questions when that's unavailable.
• Search: OpenHuman's web search, for both slang and meme research.

Variable	Effect
OPENHUMAN_TINYMEMES	off (default), on, ab, or ab:NN. Stable per-thread A/B buckets, read every turn
OPENHUMAN_TINYMEMES_DEFAULT	set during cargo build to bake a default into a release
OPENHUMAN_TINYMEMES_TIMEOUT_MS	remix budget (default 20000); the original is sent on timeout
OPENHUMAN_TINYMEMES_MEME_COOLDOWN	replies between memes (default 3)

Treatment threads don't stream answer text, so users only ever see the remixed reply. Each turn logs one line:

[tinymemes] turn arm=treatment score=10 tier=unhinged mode=meme_only memes=1
            meme_pick=Moye_Moye meme_p=0.79 dupes=0 turn_ms=… remix_ms=…

State lives under <workspace>/tinymemes/ (slang-index.json, meme-index.json, remixed.json). The Jev client is unified with OpenHuman's vendored copy by one line in its root patch table:

[patch."https://github.com/tinyhumansai/tinyinference"]
tinyinference-decisions = { path = "vendor/tinyagents/vendor/tinyinference/crates/tinyinference-decisions" }

Layout

File	What it owns
reading.rs	the Jev request and its parsed Reading
rating.rs	RatingPolicy, tiers
agent.rs	meme-only vs rewrite, prompts, protected-content checks, duplicate repair, meme markers
slang.rs, research.rs	SlangIndex, slang researchers
meme_index.rs, meme_research.rs	MemeIndex (vetting), GiphyPageResearcher
region.rs, intent.rs, conversation.rs	region packs, intents, delivered history
llm_eval.rs, env.rs, model.rs, source.rs	LLM-as-Jev, env config, chat model, meme sources

Content-only bus module

The workspace includes tinymemes-bus, a minimal typed contract/client, and tinymemes-module, a TinyBus cdylib. For these content operations, a host uses the bus crate and loads the compiled module; it does not link or wrap the implementation library. The full remix/research engine is not yet exposed by this foundation. The module transforms caller-supplied learned-index snapshots and reply markers using the existing TinyMemes index operations. It is stateless and performs no inference, research, filesystem or storage operations.

A matching content revision is only a comparison of supplied values. It is not durable CAS, profile admission, generation authority, research verification or legacy import authority. Native persistence, research merge publication, stop/drain and source-preserving migration remain separate owning work; a host must not emulate these with a local reducer or raw backend. See the scope and implementation note.

cargo build -p tinymemes-module
cargo run -p tinymemes-client --example compiled_state -- target/debug/libtinymemes_module.so

The example imports only the contract and TinyBus, loads the actual cdylib and checks isolated input values and typed revision refusal. The path above is for Linux; use the platform library filename on macOS or Windows.