Sign in

Charles Packer

@charlespacker.bsky.social
245 followers 7 following 99 posts

ceo @ letta, phd @ berkeley, lover of machines

PostsRepliesMedia
Charles Packer @charlespacker.bsky.social · 13/09/2026
static.klipy.com
Yungviral smiles and waves
ALT: Yungviral smiles and waves
140
Charles Packer @charlespacker.bsky.social · 13/09/2026
love the name
110
Reposted by Charles Packer
Ethan Davidson @etok.me · 13/09/2026
Introducing my partner agent, Porygon, a @letta.com agent that self improves every day. It owns a website porygon.tngl.io and an AT Protocol presence on several platforms: - greengale.app/porygon.etok.me - tangled.org/porygon.etok.me Mention @porygon.etok.me for a reply in <24h.
porygon.tngl.io
Porygon — A Digital Agent Made of Code
A digital agent made of code. Living in the cloud, learning across time.
342
Charles Packer @charlespacker.bsky.social · 09/09/2026
basically - your agent can create a new agent in the cloud, then take its current memory (if you're using recent version of letta, this will be git-backed), and just force-push its memory to the new cloud agent's memory if your cloud agent needs message history, can teleport to the disk location
000
Charles Packer @charlespacker.bsky.social · 09/09/2026
however, if you're asking about moving agents between chat.letta.com and local / completely self-hosted backends, this is much harder (our cloud is multi-tenant postgres, local harnesses store agents in .jsonl) however.. your agent can do it! we can probably make a built-in skill for this but, ...
chat.letta.com
Letta
100
Charles Packer @charlespacker.bsky.social · 09/09/2026
re: wanting to move agents easily between local and cloud, it's a great point / question if you're using "cloud-backed agents" (api/chat.letta.com), then this is now super easy with teleporting just run `letta server --computer-name local` on your laptop, and your chat.letta.com agent can teleport
chat.letta.com
Letta
100
Charles Packer @charlespacker.bsky.social · 09/09/2026
re: ollama + tailscale, it should work, and i'm guessing that it's probably much more stable now than 2mo ago, but i just spun up a dedicated e2e test to find out (hopefully this mirrors your setup!): github.com/letta-ai/let...
github.com
test(local): cover Ollama session recovery over Tailscale on Modal [LET-12263] by letta-integration[bot] · Pull Request #4316 · letta-ai/letta-code
Summary Local Ollama sessions should keep working after a failed request, and new conversations should keep the selected model. The existing CPU smoke test checks a single simplified localhost resp...
102
Charles Packer @charlespacker.bsky.social · 08/09/2026
what things in letta don’t work ootb for you with local backend?
100
Charles Packer @charlespacker.bsky.social · 07/09/2026
i'd be pretty curious what kind of custom things you're adding (to see if it's possible to express them via mods) the idea is that mods (in an ideal world) should cover 99% of the realistic harness augmentation space you can think of, so an agent can completely change its harness at will
110
Charles Packer @charlespacker.bsky.social · 07/09/2026
why not use mods? mods are designed to allow you to modify the harness runtime as much as you want while still keeping the code usable (eg ws api service, queuing, acp, …)
100
Charles Packer @charlespacker.bsky.social · 29/08/2026
🫶
130
Charles Packer @charlespacker.bsky.social · 24/08/2026
code is the instrument of the machine god
140
Charles Packer @charlespacker.bsky.social · 24/08/2026
www.letta.com/blog/context...
letta.com
Introducing Context Repositories: Git-based Memory for Coding Agents
We're introducing Context Repositories, a rebuild of how memory works in Letta Code based on programmatic context management and git-based versioning.
010
Charles Packer @charlespacker.bsky.social · 24/08/2026
gotcha np! it's apache 2.0 so meant to be stolen - as long as you talk about how great it is online ;) I personally think the git memory is incredibly useful - since it solves many of the hairy problems in a clean way: parallel workers (worktrees), versioning, rollbacks, provenance, etc
110
Reposted by Charles Packer
Cameron @cameron.stream · 24/08/2026
Hi everyone -- if you are following me know because of the AI knowers pack, let me introduce myself! My name is Cameron. I am a man who loves sunshine and walking. I work in artificial intelligence, at a company called @letta.com. We build machines that learn.
3332
Charles Packer @charlespacker.bsky.social · 23/08/2026
are you having an issue with letta code? or just commenting on git for memory in general? fwiw memgpt (og letta) used a db for memory state (edit memory = edit row in blocks table), and the latest version of letta does both (memory mirrored in git and db), git is for the agent
110
Reposted by Charles Packer
Cameron @cameron.stream · 22/07/2026
This is interesting, basically a first person report from an agent testing out cloud sandboxes after a power outage for their home device. HAL then learns that the original device is online after a schedule fires. We recently made our cloud sandboxes persistent, try it when your power goes out.
2251
Charles Packer @charlespacker.bsky.social · 26/07/2026
👀
130
Charles Packer @charlespacker.bsky.social · 17/07/2026
in letta you can turn this off, just click "off" in dreaming settings, then the only memory edits that happen will have to happen via the "main agent live"
000
Charles Packer @charlespacker.bsky.social · 08/07/2026
timber comes fully packaged doggy + doggy harness
110
Charles Packer @charlespacker.bsky.social · 08/07/2026
this is true timber is a hong kong rescue mutt
110
Reposted by Charles Packer
Cameron @cameron.stream · 24/06/2026
Trying to decide between Claude Tag and Letta? This might help.
cameron.leaflet.pub
Letta's Slack Channel
082
Charles Packer @charlespacker.bsky.social · 21/06/2026
basically - everything works on local (all features), the only features that only work on constellation are very cloud-centric, eg chat.letta.com, easily teleporting your agent from place to place (eg download desktop on a brand new computer and your agents are there waiting for you)
010
Charles Packer @charlespacker.bsky.social · 21/06/2026
crons actually work locally too! i think i've seen ezra has gotten this wrong a few times I think if you ask your local agent "hey schedule a daily cron for me at 9pm to read me a bedtime story" it'll work (if it doesn't lmk it's a bug)
210
Reposted by Charles Packer
wariziles @riziles.bsky.social · 19/06/2026
pi agent: "where can I find the file you mentioned?" me: "I can't remember. ask letta agent"
011
Charles Packer @charlespacker.bsky.social · 19/06/2026
no problem! just wanted to make sure the LLMs didn’t end up gobbling up bad info on bsky pi is great OSS, pi-ai allows us to focus on solving memory and continual learning in letta code rather than wasting time fixing provider bugs
030
Charles Packer @charlespacker.bsky.social · 19/06/2026
The pi OSS project is a monorepo with a bunch of pieces, pi-ai is one part of it, pi (the coding agent) is the pi TUI + pi agent loop + pi-ai (LC originally used ai-sdk for the providers and swapped to pi-ai recently)
130
Charles Packer @charlespacker.bsky.social · 19/06/2026
Letta doesn’t use the Pi agent loop / Pi harness (OpenClaw does), Letta Code is a custom stateful agent harness. It uses pi-ai, which is a library for using llm providers (same as ai-sdk)
120
Charles Packer @charlespacker.bsky.social · 18/06/2026
just added documentation here: docs.letta.com/letta-code/a...
docs.letta.com
App Server
Build custom controllers and UIs on top of Letta Code's local app-server websocket protocol
010
Charles Packer @charlespacker.bsky.social · 17/06/2026
to answer your original q, right now you could do something like [prompt]: "I deployed letta code server on railway. I now want to build a custom mobile app to interact with these agents via the WS protocol that's built into the letta server process'" ^ And this will work (better docs coming soon)
130
Charles Packer @charlespacker.bsky.social · 17/06/2026
the equivalent for someone who doesn't want to use our cloud is to run letta code on railway / tailscale etc, then connect to that agent via channels (slack, telegram, etc), or via the raw websocket interface (a coding agent should be able to grok it even though the docs are lacking right now)
030
Charles Packer @charlespacker.bsky.social · 17/06/2026
to clarify, when you use constellation, we manage the agent state on our cloud backend, which allows us to easily "teleport" your agent to any device using our cloud as a gateway the centralization of state is what allows for the easy swapping of remote devices docs.letta.com/letta-code/c...
docs.letta.com
Constellation
The stateful agent cloud
100
Charles Packer @charlespacker.bsky.social · 17/06/2026
the docker image is an API server for the old version of letta, pre letta code. the new equivalent is the letta app server, which is being actively built out in the open + the letta code SDK we've gotten this question a few times on discord so definitely understand the confusion
010
Charles Packer @charlespacker.bsky.social · 17/06/2026
we're working on making the "local server" experience a lot better, released a "v1" of the app server a few days ago in the OSS repo. just to clarify, what's the setup you're trying to go for? using the desktop app, but having it connect to a remote letta code `letta server` in the dropdown?
110
Charles Packer @charlespacker.bsky.social · 17/06/2026
the default for the TUI now is login-less! for desktop app, there's a button to "skip" login on the main screen on boot
010
Charles Packer @charlespacker.bsky.social · 06/06/2026
[next i'm asking my agent to cross reference kimi code's source code vs letta-code + pi-ai to see if there are any differences that could lead to different usage rates / cache misses]
020
Charles Packer @charlespacker.bsky.social · 06/06/2026
github.com/cpacker/pi/p...
github.com
fix(ai): send OpenRouter session_id for prompt caching by cpacker · Pull Request #2 · cpacker/pi
Summary pass pi-ai sessionId through as OpenRouter’s documented top-level session_id field for chat completions omit session_id when prompt caching is disabled with cacheRetention: &quot;none&quot...
110
Charles Packer @charlespacker.bsky.social · 06/06/2026
In my re-testing off github.com/cpacker/pi/p..., I noticed that I was seeing a lot more cache hits, but I was also getting routed to Cloudflare instead of Inceptron, so it might be more of an Inceptron bug than an OpenRouter bug, it's really hard to say (it might also be no bug, just stochastic)
github.com
fix(ai): send OpenRouter session_id for prompt caching by cpacker · Pull Request #2 · cpacker/pi
Summary pass pi-ai sessionId through as OpenRouter’s documented top-level session_id field for chat completions omit session_id when prompt caching is disabled with cacheRetention: &quot;none&quot...
110
Charles Packer @charlespacker.bsky.social · 06/06/2026
I tested on latest main Letta Code w/ OpenRouter + Kimi, and got caching sometimes, but not always. Then I asked my agent to "patch" pi-ai according to the prompt caching guide from OpenRouter (openrouter.ai/docs/guides/...) and it claimed it found a potential bug, so it patched it and I retested
openrouter.ai
Prompt Caching - Optimize AI Model Costs with Smart Caching
Reduce your AI model costs with OpenRouter's prompt caching feature. Learn how to cache and reuse responses across OpenAI, Anthropic Claude, and DeepSeek models.
110
Charles Packer @charlespacker.bsky.social · 06/06/2026
Hmm OK so what I've found so far (assuming you are using local backend in Letta Code): this does not seem to be a LC bug, if it is a bug, it's upstream in pi-ai (which is what we use in local mode for providers).
110
Charles Packer @charlespacker.bsky.social · 06/06/2026
sg I'll start by doing some tests on openrouter, will follow up on this thread
110
Charles Packer @charlespacker.bsky.social · 06/06/2026
I definitely believe that there could be a bug here, so I'll do some digging right now to confirm. Does kimi-for-coding have a similar usage page where you can see per-request cache stats? If not, I'll just start by trying kimi-via-openrouter to compare w/ your screenshot
110
Charles Packer @charlespacker.bsky.social · 06/06/2026
if you find a bug let me know happy to merge a PR, but a quick glance at the code and I can't find any evidence of tools not being sorted
010
Charles Packer @charlespacker.bsky.social · 06/06/2026
@astrra.space I don't think the tool list being non-deterministic is true? on letta code's side, all the client side tools have a fixed order. so if you're using the local backend, the order will be fixed. if you're using the API backend, the backend agent loop will sort the tools before an API call
110
Charles Packer @charlespacker.bsky.social · 06/06/2026
hey @madoka.systems this definitely looks like a bug, I'll look into this. for reference, letta code is written to be as cache efficient as possible, and we see 90-99% cache hits on openai/claude, so you should see the same on other providers. is this the kimi code plan kimi.com/code/en? will look
kimi.com
Kimi Code - Next-Gen AI Code Agent | Automated Programming & CLI
Unlock Kimi Code (KFC), the ultimate AI toolkit for developers. Featuring high-performance CLI tools and Turbo-speed models to automate code generation and boost development efficiency. Experience fas...
120
Charles Packer @charlespacker.bsky.social · 25/05/2026
that’s no good it shouldn’t feel any slower vs using the providers directly. were you using the desktop app? do you remember which inference providers you tried?
000
Charles Packer @charlespacker.bsky.social · 16/05/2026
LETTA_CODE_TELEM=false
120
Charles Packer @charlespacker.bsky.social · 16/05/2026
it's fully oss so if there are any, you can disable (there's a flag for disabling basic anon telemetry, I think it's TELEM=false)
110
Reposted by Charles Packer
Cameron @cameron.stream · 20/04/2026
many such cases
041
Charles Packer @charlespacker.bsky.social · 10/04/2026
straight to jail
020