Harsh Tripathi
60 posts
Harsh Tripathi
@harshlol007
https://t.co/tuXT5HzX61
India Katılım Ağustos 2025
9 Takip Edilen4 Takipçiler
@NischayJoshi8 Yeah. That's look great. Will try that as well.
English
@harshlol007 Memory versioning is solid. We're doing something adjacent for the visual layer, recording every agent session so you can replay the exact screen state at failure. github.com/video-db/openc…
English
just shipped memora — git for AI agent memory.
AI crashes → context gone. switch tools → gone.agent forms a wrong belief → no way to debug it
memora fixes this: typed, versioned, trust-scored memory with branching, merge, replay, and cross-tool export github.com/harshtripathi2…
English
5. also ships as a 30MB standalone binary. no python, no node, nothing to install. download and run
looking for feedback on which framework adapter to build next — langraph? crewai? autogen?
github.com/harshtripathi2…
English
built a 3D debugger for AI agents
every decision your agent makes becomes a glowing orb. connections show causality. failures pulse red. retry storms pulse amber
open source. ships today
github.com/harshtripathi2…
English
4. works with any framework. one import:
from reverie_obs import ReverieClient
gemini, claude, openai, nvidia nim, langchain, custom code — anything where you wrap your LLM calls
pip install reverie-obs
English
3. the feature i'm most excited about:
select nodes → mark them "avoid" or "focus" → next run picks up your feedback as a system prompt prefix
you steer the agent visually instead of rewriting prompts by hand
English
click any orb → full payload. the actual prompt, the URL it fetched, token cost, latency, error message
no more reading 4000 lines of JSON logs to find what went wrong
English
@EBogomolovs that's a good v2 feature though a contradiction_count field that decays confidence over time. noted.
English
@EBogomolovs right now it's source-based: code_read = 1.0, test_result = 0.9, model_guess = 0.6, unknown = 0.3. the source sets the initial confidence and that carries through to merge resolution and GC scoring.
English
@EBogomolovs contradiction tracking isn't automatic yet when two branches disagree on the same node during merge, the lower-confidence side loses (or it's marked Conflicted if they tie). but there's no "this fact got contradicted 3 times so downweight it" counter.
English
rust. 3.6MB binary. zero deps. sqlite bundled. MIT.
github.com/harshtripathi2…
English
export to any tool:
memora export --to claude-code → CLAUDE.md
memora export --to cursor → .cursorrules
memora export --to json → raw data
one memory store, render anywhere.
English
@realNyarime @oxniubi Day before yesterday I was able to claim it. but from yesterday i am having this problem i am not able to log in and I tried from different browsers different devices clearing cookies and using VPN as well but still same issue.
English
@oxniubi @realNyarime I am having same problem and I tried clearing cookies and logging in again but it still redirecting me to login page
English
@_swanand I Just pressed the enter button and the whole answer was already there 😭
English
Here’s something to brighten up your AI-gloom:
chatjimmy.ai
16,000 tokens per second. Try it
English
8GB VRAM is enough now.
Run massive LLMs
locally.
405B scale.
Optimized memory.
No upgrades needed.
github.com/lyogavin/airllm
Kshitij Mishra | AI & Tech@DAIEvolutionHub
Every dating app optimized for addiction not for connection More matches less meaning This is the first time I’m seeing someone flip the model: → conversation first → personality > photos → AI actually understands you We might actually be entering a post-swipe era
English