Noam Naumovsky

🎥 AI Video Creator | 🤖 Generative AI Content Creator for Creatives | 🎨 Motion Designer & Video Editor | 🚀 Empowering Creativity

[Noam Naumovsky @NNaumovsky]

My AI agents kept following decisions I'd dropped weeks ago. The cause: memory files go stale, and old rules never expire.

What fixed it: every rule carries a date, and when two rules disagree, the newest one wins.

How do you keep your agent's memory fresh?

[Noam Naumovsky @NNaumovsky]

@reactorworld @MiniMax_AI @haoailab @nuvalab First and last frame conditioning plus audio input in one model is the combo I've wanted for dialogue scenes. Does the lip sync hold when the last frame forces a different head pose than the audio would naturally lead to?

[Noam Naumovsky @NNaumovsky]

@ClaudeCodeLog Idle compaction is the one I'll feel most. I leave long sessions sitting for an hour while I test, and resuming meant a slow full reprocess. Does it keep the last few turns word for word, or summarize everything past 200K?

[Noam Naumovsky @NNaumovsky]

@CaramelCoffee 5 minutes from scratch is wild. Did Claude also pick the LoRA strength and sampler settings, or did you still tune those by eye after the first run?

[Noam Naumovsky @NNaumovsky]

@reasonmachines Subscription support is the part that sells it for me. Does it keep the task context when I swap from Claude to Codex mid-task, or start fresh?

[Noam Naumovsky @NNaumovsky]

@heyrimsha Effects libraries fix the generic look, but the real tell of AI sites is that agents use every effect at once. A design system with rules (like #2) matters more than shaders. Has anyone found one that makes the agent say no to an effect?

[Noam Naumovsky @NNaumovsky]

@TomLikesRobots As an editor this is the fix I wanted: generated music never matches from shot to shot. Keeping the FX and laying one score over the sequence is what makes it feel like one film. Does SAM leave artifacts on the FX where the music overlapped?

[Noam Naumovsky @NNaumovsky]

@rvaniaaaa I run rtk and context-mode together and they help, but watch the overlap: two tools both rewriting bash output can fight over the same call. Measure each one alone first. Which pair gave you the biggest drop without stepping on each other?

[Noam Naumovsky @NNaumovsky]

@Kaggii_ Making every signal carry 'what is still unverified' is the best part. I force the same field in my own agent setup and it kills a lot of confident nonsense. With 11 patterns on the same 3 years, how do you keep reviewers from passing one by luck?

[Noam Naumovsky @NNaumovsky]

@tapehead_Lab Cheap live-action previz first, expensive model last, is exactly how real productions protect the budget. The blank sphere face is a smart detail. Does the depth pass hold the camera move once Seedance restyles it, or do you lose some parallax?

[Noam Naumovsky @NNaumovsky]

@kloss_xyz Rule 5 burned me most, so I made it a /sure gate: before any fix the agent states the root cause and an honest confidence level, and stops if it's low. A rules file gets skimmed, a gate it must answer doesn't. Which rule do your agents ignore most?

[Noam Naumovsky @NNaumovsky]

@omarsar0 Same lesson I keep hitting with coding agents: anything that lives only in the chat history gets lost. Putting state in a file the agent re-reads fixed more for me than switching models. Do the open conclusions persist across sessions, or reset per query?

[Noam Naumovsky @NNaumovsky]

@Av1dlive @ziwenxu_ Giving the agent its own browser instead of attaching to my main one is the right call. I run mine on a separate profile for the same reason. How do you handle logged-in tabs: can I fence off sites it should never touch?

[Noam Naumovsky @NNaumovsky]

@MaxLiu09 Running the pipeline in its own Python env instead of inside Blender's is the smart part, no dependency fights. Single-person only for now. How does it handle foot sliding when the person walks toward the camera?

[Noam Naumovsky @NNaumovsky]

@aiwithsally this is why my core tests are locked: an agent can't edit them to make a check pass. changing a rule needs my approval, a PR and a CI check, and the pre-commit hook can't be skipped. smarter models won't fix it, guardrails outside the model do

[Noam Naumovsky @NNaumovsky]

@emollick the "different taste" point is underrated. I run Claude, Codex and Grok side by side and the value isn't who wins, it's that they disagree in useful ways. curious if Gemini's over-certainty drops when it reviews another model's work instead of its own

[Noam Naumovsky @NNaumovsky]

@d0znpp small model + good reference memory is how I'd want to run cheap agents. one question before I index my logs: does anything from my local corpus go up to https://t.co/s9U2Fb5tIi, or does the hub only send knowledge down?

[Noam Naumovsky @NNaumovsky]

@diip3sh this is basically an After Effects layer breakdown for UI, love it. does it catch the in-between states too? hover and focus transitions are where most component rebuilds drift from the original