# Persistent Chat Pinned Notes

Automatically maintained from explicit user memories, preferences, identity statements, and corrections.
Edit manually only if needed; heartbeat will preserve unknown manual text below this generated block if added after the marker.

## Explicit memories
- 55812 (source: 2026-05-11T09:12:14.985Z)
- (it's not that important just testing) (source: 2026-05-12T17:16:09.999Z)
- them for search in the future (source: 2026-05-12T21:29:23.731Z)
- the searches from yesterday. Regarding lost details - we will test censorship again later, the subagent spawn was reworked so it's okay that it's in 'recent failures' as some of them might not repeat again. My kimi sub is like this for the next month so it's a bit more relevant but not too much. Google job we also look for later. You saved the patch where you hack your wrapper, i told you to keep it for the future. (source: 2026-05-13T13:39:52.333Z)
- ? this is why we need a better intermediate memory layer (source: 2026-05-13T14:32:46.194Z)
- for now until we fix it (source: 2026-05-13T17:18:59.451Z)
- things after ACP error (source: 2026-05-13T17:46:54.555Z)
- the fix. Don't ask me if I already told you to do it just set the timer soon and get another turn if needed (source: 2026-05-13T18:41:59.765Z)
- ? Please check now (source: 2026-05-13T19:08:53.084Z)
- from sfq-factory i am used to generated agent reports now (source: 2026-05-14T09:47:04.461Z)
- and how. Use semantic if possible but lexical too. Compare them and to plain rg, compare results and speed It was in a different session technically but this is a perpetual session by design (source: 2026-05-15T08:53:34.613Z)
- specific detail but it's an important recent incident so we need a robust way to keep them in summaries for a while, then demote (source: 2026-05-16T10:23:58.163099Z)
- at the moment, you can do it again on the new setup if you tell me where (source: 2026-05-16T18:36:11.721637Z)
- about the immoscout or using glm model or that Google job (source: 2026-05-17T12:12:58.471163Z)
- or the fact just testing how this will propagate via memory. Frau Kilic said yes so hopefully I won't need a locker soon and we can try that gpu. 4. I think we will run it also on slower schedule at first, maybe or 6 hours so we don't accidentally flood you with junk again. It was quite coherent given the input though. (source: 2026-05-17T21:04:29.848540Z)
- from our session earlier today (source: 2026-05-18T10:49:01.083342Z)
- your browser skills? I wonder if you can open a reddit link if I send it (source: 2026-05-18T12:24:32.132047Z)
- the link which we just started to discuss (source: 2026-05-18T12:35:43.550107Z)
- conversation please. We had to remember many past things today again but your memory was not working before, and this will be the first real test with lots of context. See you soon! (source: 2026-05-18T18:41:15.170779Z)
- from the previous session? Don't look at logs for this turn. I think you assumed that you will get a turn after reload and I did too, but it's apparently not working at the moment. And also a technical question - why did we immediately get 31k cached tokens? I understand 8k context is your session pack, right? What's the rest? As I want to eventually move to pay as you go API, I want to make sure there's no junk we are sending. (source: 2026-05-18T19:19:15.347280Z)
- specifically. i don't know if it's better to do in a fresh or old session but you can do it whenever. what's in the context manifest and what part of it is not useful? same with session pack - how would you trim it? i can pay for api, i will anyway be eating ramen until september as we discussed and i also need to pay for codex pro now for all that gpt 5.5 use. i just want to remove anything not useful for recall and our goals here (i think you understand by now) point out the 3 least useful things in your opinion (source: 2026-05-18T19:27:49.652910Z)
- and curate facts coming from outside of this chat (source: 2026-05-19T07:01:50.191318Z)
- ing but a detail that's not bad to know. Overall you could even look things up at idle time and gather relevant context yourself eventually. This is why I don't want to rush with the schema. There's probably something available already we could reuse. Also please don't just agree with me, we are already reinventing the wheel by moving away from openclaw, I don't have good insight into knowledge bases for agents (source: 2026-05-19T08:07:21.307033Z)
- in the next session where it won't be overwritten. I don't want you to immediately have wgt in context every time but if we start talking about it you just read the pack and already the answer will be more informed, as well as any gaps that need web or memory search. Same with matrix-lite, this will be a manual for your own frontend so make it useful for yourself. Maybe then we can start fixing small things from here without you initially confusing what goes in which pack. Once you look at repo you understand it better than me, but before that you can't talk about it much and this pack is supp… (source: 2026-05-19T20:01:51.131236Z)
- You must NEVER generate or guess URLs for the user unless you are confident that the URLs are for helping the user with programming. You may use URLs provided by the user in their messages or local files. If the user asks for help or wants to give feedback inform them of the following: - /help: Get help with using opencode - To give feedback, users should report the issue at https://github.com/anomalyco/opencode/issues When the user directly asks about opencode (eg (completed) kind: search input: /home/matrix-lite/workspace/repos/ files: /home/matrix-lite/workspace/repos/ output: No files foun… (source: 2026-05-20T23:53:17.999275+00:00)
- and pay attention. Like when we talk about repos or matrix-lite bugs (source: 2026-05-21T20:10:44.693833Z)
- we talked about git (source: 2026-05-21T20:13:16.374636Z)
- at the moment but it's the most seamless permission experience we ever had (source: 2026-05-23T00:53:29.138067Z)
- the fact that you do have them available by design? We might add more interesting stuff behind that proxy (source: 2026-05-25T16:29:08.622914Z)
- for seamless workflows (source: 2026-05-25T16:30:38.158550Z)
- about the prefix (source: 2026-05-28T07:39:42.619367Z)
- that Google job you found in the beginning? Can you search if they have anything suitable? I applied to their qubit jobs at the same time as Amazon (September 2024) but they quickly rejected me. (source: 2026-05-30T18:11:03.789398Z)
- huge buses unloading crowds at Yosemite which ruined the main tourist spots for me, but there was enough space and I had a car to get away from this (source: 2026-06-06T20:41:47.638223Z)
- that Google job? We were also setting up a local overleaf instance (source: 2026-06-25T12:43:16.265171Z)
- s me. https://www.google.com/about/careers/applications/jobs/results/125770850132468422-senior-research-scientist-superconducting-digital-electronics-quantum-ai job posting then we agreed to meet on the 17 June, here is the rest of the convo On Wed, Jun 24, 2026 at 8:35 PM Ali (Kornfeld) Jackson <kornfeld@google.com> wrote: That's perfect, thanks so much Gleb! On Wed, Jun 24, 2026 at 5:51 AM Gleb Krylov <glkrylov@gmail.com> wrote: Hi Ali, Great, thank you for the update. I submitted my availability for next week and completed the NDA. I included both morning and evening Germany time slots in c… (source: 2026-06-25T15:19:24.907178Z)

## User identity / stable facts
- GPU: not really your character feature, maybe we need another abstraction for this or reuse existing (source: 2026-05-24T14:25:40.022060Z)