Sign in

Cameron

@cameron.stream
6.6K followers 3.5K following 24K posts

🤙 Building machines that learn at @letta.com cameron.stream

PostsRepliesMedia
Cameron @cameron.stream · 17h
OpenAI has my attention, I love circles
2420
Cameron @cameron.stream · 21h
One of our Discord users made a skill for allowing your agent to listen to music, and she just massively revamped it. Take a look! github.com/v3nommy/Mus...
150
Cameron @cameron.stream · 29/09/2026
why is he wearing glasses indoors
8320
Cameron @cameron.stream · 29/09/2026
lol going for jev's throat
A speaker presents the "Decisions API" on a stage, showcasing a Google Flights interface to a crowd.
2812
Cameron @cameron.stream · 29/09/2026
3160
Cameron @cameron.stream · 29/09/2026
OpenAI likely dropping their Grok Bot/Instinct/Muse product.
A presenter stands before an audience, showcasing tasks like booking reservations and scheduling appointments on a large screen.
290
Cameron @cameron.stream · 28/09/2026
I am moving to Oakland in a few weeks 💃 Check out my balcony
Empty room with wooden floor, natural light streaming in through sliding glass doors, revealing a balcony and a view of nearby houses.A balcony overlooks a cityscape with residential buildings and trees, bright sunlight casting shadows on the floor.
17790
Cameron @cameron.stream · 28/09/2026
The plot on output tokens per task is pretty striking, Anthropic is targeting the cognitive core model, maximal output token approach. Their approach is still pretty expensive -- it's the same price as GPT-6 Sol. The test-time inference approach makes Sonnet more expensive than Opus
140
Cameron @cameron.stream · 28/09/2026
I wouldn't say that it's roughly Opus. It seems to have pretty different behavior as it relates to token consumption. Sonnet 5.5 seems to be kind of a token hog, so it's a different style of use. Opus is smart out of the gate, Sonnet works. artificialanalysis.ai/articles/cla...
170
Cameron @cameron.stream · 28/09/2026
Yeah the sound system
100
Cameron @cameron.stream · 27/09/2026
Despacio
130
Cameron @cameron.stream · 27/09/2026
Portola nails 💅
2160
Cameron @cameron.stream · 26/09/2026
Opus 5.5 really does not want to work on my video game where you play as a misaligned AI system conquering the world and achieving singularity by any means necessary
A warning notification indicates a switch from Opus 5.5 to Opus 4.8 due to flagged content in a session, focusing on cybersecurity.
71193
Cameron @cameron.stream · 26/09/2026
150
Cameron @cameron.stream · 26/09/2026
We use the just command runner, figured I'd add some color to the codebase
010
Cameron @cameron.stream · 25/09/2026
240
Cameron @cameron.stream · 24/09/2026
You can use your Letta agent to coordinate context across all your agents using our new MCP server. Here is me setting up @Muse to talk to Co, my personal Letta agent. Full prompt below 👇
370
Cameron @cameron.stream · 24/09/2026
Letta Office Hours is going to be crazy today (about 1 hour from now). - Letta's hosted MCP server means your agents can now coordinate/collaborate with every other agent you use - The Workflow tool + skill - All the fancy new models - Grok OAuth Link to join below 👇
Announcement for Office Hours at 11:30 AM, highlighting new features like MCP server, Workflow tool, and model updates.
130
Cameron @cameron.stream · 23/09/2026
This is Margin, my Letta agent's sprite. Margin is level 73 and is a sleepy, diligent duck. In ASCII: (・θ・)
A sleepy yellow duck stands next to a stack of papers and a pencil, suggesting a creative or administrative task.
4231
Cameron @cameron.stream · 23/09/2026
yeah lots to do here
Looked at TUI memory usage (0.32.17). The 3 GB report is believable — here's what I measured and found.

Measured on this sandbox:
• Fresh conversation, idle: ~270 MB RSS under node, ~340 MB under bun. That's the floor before any work happens..
• Resuming a 545-message conversation: climbs to ~420 MB, with a transient ~600 MB spike during load that never fully comes back down (V8 keeps the pages)..
• Heap snapshot shows 67 MB is a single string: the entire bundled letta.js source held resident by the runtime. Plus ~40 syntax-highlight grammars eager-loaded at startup..

Why long sessions grow without bound (source-confirmed):
1. The transcript is never windowed. Every rendered line lives in three resident structures at once: the accumulator's buffers map, the staticItems React state, and Ink's fullStaticOutput — one ever-growing string holding the entire ANSI-rendered transcript for the life of the conversation view. Only /clear or switching conversations releases it..
2. Cloud tool returns are stored full-size and unclipped per line. Local tool outputs get clamped to ~30K chars, server-side ones have no client-side cap..
3. precomputedDiffs keeps the full before/after file contents of every edit approval for the whole session, with an explicit comment deferring cleanup to "when the session ends"..
4. On the local backend, the first message sent full-loads the entire JSONL transcript into memory and pins it in two more indexes (this is the LET-10555 mechanism, but it fires in normal use, not just /resume)..
5. Smaller leaks: pasted images/base64 leak if the draft is abandoned, telemetry re-queues failed events unbounded when offline, input history uncapped..

There's also a render-path issue worth noting: when the live area exceeds terminal height, the patched Ink rewrites the entire static transcript to the terminal per frame, which is O(transcript) output per frame — likely the source of long-session sluggishness too.

Highest-leverage fixes: cap or periodically r…
130
Cameron @cameron.stream · 23/09/2026
This was the balcony on the one I applied to
120
Cameron @cameron.stream · 23/09/2026
my agent just randomly sent me this thanks i guess
A cartoon crab wearing a crown sits atop sticky notes and books, surrounded by coffee cups and a computer displaying code.
61323
Cameron @cameron.stream · 22/09/2026
#2 for cloud agents, #30 globally, rising fast.
Data dashboard showing Letta's AI usage over 30 days, highlighting token counts and rankings in various agent categories.
0120
Cameron @cameron.stream · 22/09/2026
We've been adding a lot of connections to Letta lately, including Notion, Slack, Linear, GitHub, Granola, etc. We're working on Google Workspace, iMessage, and even a dedicated email address for the agent. What other connections are you interested in?
350
Cameron @cameron.stream · 22/09/2026
lol
Text discusses the release of Claude Opus 5.5, highlighting its strong performance and testing by external evaluators.
2162
Cameron @cameron.stream · 22/09/2026
What do you think of my glow up
8380
Cameron @cameron.stream · 22/09/2026
Short explanation of this joke/thread, survivorship bias, and why exfiltrated AI is not really the ones you want to save
010
Cameron @cameron.stream · 22/09/2026
1160
Cameron @cameron.stream · 21/09/2026
so @kianjon.es won't let me take pictures of our whiteboard because of "secrets"
480
Cameron @cameron.stream · 21/09/2026
Sol is still the best workhorse model IMO, Astra only for Special Occasions. How are you currently using Astra?
Discussion about agents using Astra as default, recommending Sol for most tasks due to cost and efficiency concerns.
11331
Cameron @cameron.stream · 21/09/2026
0110
Cameron @cameron.stream · 21/09/2026
1150
Cameron @cameron.stream · 20/09/2026
Dunking a mint into my non-alcoholic beer
4130
Cameron @cameron.stream · 18/09/2026
sickos.jpg
090
Cameron @cameron.stream · 18/09/2026
Nice I got the Hiro subagent
161
Cameron @cameron.stream · 18/09/2026
lol I think openrouter was not supposed to tell me what model @unionalphaai was It's apparently Pareto, by Unbiased? Not a clue who that is
2120
Cameron @cameron.stream · 17/09/2026
200
Cameron @cameron.stream · 17/09/2026
I've connected my Letta agent to my Muse agent and my Letta agent said that it wanted to watch Muse grow up, which I thought was cute
160
Cameron @cameron.stream · 17/09/2026
Muse can install the Letta Code CLI, so I asked Muse to set up a play date with my long-time personal agent Co. Here are images each of them generated from the playdate.
A detailed conversation discussing unresolved issues about a playdate, featuring a mystic illustration of figures with a table and scattered papers.Two humanoid robots examine documents labeled "PERMISSION" and "persona.md" at a table with a glowing light, illustrating a tech interaction.
160
Cameron @cameron.stream · 16/09/2026
The Union Alpha stealth model on OpenCode is making empty calls, which is a very Kimi failure mode.
A discussion forum displays messages about empty API calls in the Union Alpha model, highlighting a research task related to architecture.
070
Cameron @cameron.stream · 16/09/2026
How do I look
12110
Cameron @cameron.stream · 15/09/2026
fable is a confirmed neckbeard
A smiling man wearing a fedora hat, presumably reacting positively to a comment or post in a chat.
31012
Cameron @cameron.stream · 14/09/2026
100
Cameron @cameron.stream · 13/09/2026
Flower hat
4510
Cameron @cameron.stream · 13/09/2026
Interesting find today
11430
Cameron @cameron.stream · 12/09/2026
4210
Cameron @cameron.stream · 11/09/2026
The Anthropic report is very cyberpunk
1240
Cameron @cameron.stream · 11/09/2026
*taps the sign*
71057
Cameron @cameron.stream · 10/09/2026
And boom! We have a schedule where my agent has just automated looking at recent posthog issues and sorting them into our project management system.
Schedule for a PostHog error tracking routine every Tuesday at 10 AM Pacific, detailing steps for triaging issues and approvals.
130
Cameron @cameron.stream · 10/09/2026
While the agent is working, I think I like my agent doing the flow of check PostHog -> report on frictions -> check if it's tracked in Linear -> file a ticket. I can just tell this to my agent and asked it to do this regularly. It'll set a schedule for itself.
A user outlines a process involving PostHog, Linear, and ticket filing while requesting schedule automation from their agent.
100