Dheeraj Manwani

Full Stack Developer | React • Node • Next.js

[Dheeraj Manwani @dkManwani2000]

Here's how I handled "Context Rot" in Tau (𝜏)

⚡ Tier 1 : Compaction (kicks in at 60% context usage)
Shrinks tool outputs. Collapses already-read files. Trims bloated results. No LLM needed, no data lost.

🧠 Tier 2 : Summarization (kicks in at 75% if compaction isn't enough)
Recent messages stay exactly as they are. Compresses older history into a tight running summary. Persists to DB - survives crashes and reloads.

🎯 Token counting that self-corrects.
Every turn, the API's real token count are considered. No fixed guesses - thresholds stay accurate as the conversation grows.

Two tiers. Zero manual intervention. Context stays clean across long multi-step builds ⚒️

[Dheeraj Manwani @dkManwani2000]

Here's how I handled "Context Rot" in Tau (𝜏)

⚡ Tier 1 : Compaction (kicks in at 60% context usage)
Shrinks tool outputs. Collapses already-read files. Trims bloated results. No LLM needed, no data lost.

🧠 Tier 2 : Summarization (kicks in at 75% if compaction isn't enough)
Recent messages stay exactly as they are. Compresses older history into a tight running summary. Persists to DB - survives crashes and reloads.

🎯 Token counting that self-corrects.
Every turn, the API's real token count are considered. No fixed guesses - thresholds stay accurate as the conversation grows.

Two tiers. Zero manual intervention. Context stays clean across long multi-step builds ⚒️

[Dheeraj Manwani @dkManwani2000]

Ever noticed your AI giving brilliant answers at the start… and garbage by the end?
You didn't imagine it. There's a name for it - Context Rot 🗑️

No error. No warning. Just an AI quietly losing the plot
- contradicting itself, forgetting context, hallucinating and still sounding confident doing it.

Every LLM has a fixed window of what it can "see." As it fills up, the important stuff gets buried. The model doesn't get dumber - it just stops seeing clearly.

The attached graph shows when and how fast quality drops as the context window fills up 📉 1/2

[Dheeraj Manwani @dkManwani2000]

@Ribhu_bh It depends upon the type of app. But, In general you should preserve some recent conversations and some important entities like names, goals, decisions that model will likely need again

[Dheeraj Manwani @dkManwani2000]

Ever noticed your AI giving brilliant answers at the start… and garbage by the end?
You didn't imagine it. There's a name for it - Context Rot 🗑️

No error. No warning. Just an AI quietly losing the plot
- contradicting itself, forgetting context, hallucinating and still sounding confident doing it.

Every LLM has a fixed window of what it can "see." As it fills up, the important stuff gets buried. The model doesn't get dumber - it just stops seeing clearly.

The attached graph shows when and how fast quality drops as the context window fills up 📉 1/2

[Dheeraj Manwani @dkManwani2000]

@pandeygsundaram Thanks 😊

[Dheeraj Manwani @dkManwani2000]

Been deep in agentic loops and AI coding harness architecture for 10 days.
Built v1 of Tau (τ) a full-stack builder that takes a prompt and ships frontend, backend with DB autonomously ⚡️
It plans → breaks down tasks → codes → spawns sub-agents for debugging or exploration → searches the web when needed.
Any feedback / suggestion is appreciated 🚀 @buildinpublic @kirat_tw
Demo 👇

[Dheeraj Manwani @dkManwani2000]

Code: https://t.co/yvW5FG4Y0w

[Dheeraj Manwani @dkManwani2000]

Been deep in agentic loops and AI coding harness architecture for 10 days.
Built v1 of Tau (τ) a full-stack builder that takes a prompt and ships frontend, backend with DB autonomously ⚡️
It plans → breaks down tasks → codes → spawns sub-agents for debugging or exploration → searches the web when needed.
Any feedback / suggestion is appreciated 🚀 @buildinpublic @kirat_tw
Demo 👇

[Dheeraj Manwani @dkManwani2000]

@usman_akinyemi Thanks 😊

[adil @whoisasx]

been wanting to understand matching engines from first principles.

so i built one in c++.
repo: https://t.co/RbBzcbhU4m

nothing crazy.

1 market = 1 ordered worker.
N markets = N parallel workers.
inside a market, commands stay sequential. deterministic order book.

across markets, everything scales independently.

redpanda → parse → validate → dedupe → match → update → emit → checkpoint

i was never trying to build the fastest engine on earth.

i wanted something i could understand completely, replay, benchmark and keep extending.

turns out simple systems teach you the most.

would love feedback from backend/systems/trading infra people.

@kartikkktomarrr thank you for that talk.

[Dheeraj Manwani @dkManwani2000]

LLMs do not store data, we need to somehow explain them what is our goal and what we are talking about currently. When we provide data like conversation history (between the user and LLM) and key findings (like name, age etc) this is called memory. Not everything is worth sharing with the LLM, so deciding the amount of data to be shared, the way to share data, is known as Context Engineering (There is so much more to context engineering, this definition is just to simplify it).

[Dheeraj Manwani @dkManwani2000]

Thank you @Samiksha2908 for a wonderful and informative session on context engineering and memory.

[Dheeraj Manwani @dkManwani2000]

Session 2 on memory - Multimodal memory and real world usage of memory by platforms like Openclaw, Claude code, Codex 🚀 @buildinpublic

[Dheeraj Manwani @dkManwani2000]

Session 2 on memory - Multimodal memory and real world usage of memory by platforms like Openclaw, Claude code, Codex 🚀 @buildinpublic

[Dheeraj Manwani @dkManwani2000]

@10xdhruv @Samiksha2908 @kirat_tw @100xDevs 🤝

[Dheeraj Manwani @dkManwani2000]

Latent at @100xSchool 🙌

[Dheeraj Manwani @dkManwani2000]

@harivansh666 @Samiksha2908 @kirat_tw @100xDevs Yes

[Dheeraj Manwani @dkManwani2000]

Thank you @Samiksha2908 for a wonderful and informative session on context engineering and memory.

[Dheeraj Manwani @dkManwani2000]

Latent at @100xSchool 🙌