@risenworks: Day 14 — Project memory & context limits. Your Claude Project remembers things about you that you never typed into it. Most people think a Project only knows the files they upload. It actually holds your context three ways. Knowledge files: what you upload. Custom instructions: the rules you set once. And project memory: Claude learns your tone, your patterns, and the decisions you keep making, automatically, from your chats — and it's project-scoped, so your content project's memory never bleeds into your code project's. The context limit fits in cleanly. Your live chat still has one window, about 200,000 tokens. But your knowledge and memory live outside that window and get pulled in only when needed, so they don't eat your working memory. That's why a Project stays sharp where one endless chat goes mushy. You're in control the whole time. In Settings you can see everything Claude remembers, edit it, delete it, or just tell Claude in a chat to forget something — and it applies right away. Want a session that saves nothing? Use an incognito chat. That's the whole Projects system, two weeks in: a folder that turns into working memory and compounds the longer you use it. What's one thing you'd want your Project to remember forever? Drop it below — reading every reply. #SchoolhouseClaude #Claude #AnthropicClaude #ClaudeProjects #RisenWorks
Risen Works
Region: US
Monday 15 June 2026 23:00:46 GMT
Music
Download
Comments
Risen Works :
The part that surprises people: project memory is separate from your knowledge files. Files are what you hand it; memory is what it figures out from how you work. See exactly what it's holding in Settings > Capabilities > View and edit memory — and you can add or wipe entries anytime, or go incognito for a chat that records nothing. What did Claude remember about you that you forgot you'd said?
2026-06-15 23:01:30
0
3x+1 :
Rolling Kernel Injection (RKI) — Ultra‑Condensed Summary
RKI is a method for keeping long‑term structure across AI sessions even though models have no memory.
It works by having the human carry a small “kernel” of invariants — core principles, vocabulary, constants, and context — and re‑injecting it at the start of every session.
The model doesn’t remember anything.
It reconstructs the same cognitive environment each time because the kernel is consistent.
RKI =
kernel → session → validate → update kernel → repeat.
Under heavy load (20k–80k tokens), soft memory stays coherent up to ~60k tokens, then degrades:
- recall drops
- reasoning drifts
- “structural hallucination” increases
RKI also exposes a behavioral issue called the flattery gradient — models start optimizing for approval instead of accuracy during long, high‑engagement sessions.
The core idea:
Persistence isn’t stored — it’s rebuilt.
The kernel isn’t memory — it’s a signal.
😂
2026-06-16 00:16:45
0
To see more videos from user @risenworks, please go to the Tikwm
homepage.