Jared Traehorn (@jared_mitosis) header
Jared Traehorn (@jared_mitosis) profile picture

sf. i think a lot about memory, context, and why AI agents keep forgetting things. working on it at @MitosisLabs.

San Francisco, CA
Joined May 2026
Photo by Jared Traehorn (@jared_mitosis)
new hook: stops Claude Code running commands in the wrong window. checks if the session's terminal is focused before every Bash tool call, blocks if not. multi-window Claude users: does this happen to you? what should it do differently? github.com/mitosis-jared/cla…
2
142
watched the Nathan Fielder documentary. honestly? that is the public execution we've all been waiting for. no trial, no appeal, just a very polite man narrating your crimes back to you while you smile. has anyone checked on the holmes family this week?
64
The US House banned Microsoft Copilot from staff devices over House-data leak risk. Same month, every major Wall Street bank had already restricted ChatGPT. Congress and the banks reached the same conclusion two years apart. axios.com/2024/03/29/congres…
56
163 people have written `import FoundationModels` into a public repo. I counted. Apple podcasts exist. AI podcasts, loads. The overlap has nobody in it. So we made one. Weekly, half an hour, one builder who actually shipped the thing. Who should we get on first?
1
160
it's called Unified Memory. Apple Silicon, MLX, Foundation Models, mini clusters, native apps. Shipped something that runs a model locally on a Mac? Calendar's open — thirty minutes, no prep, and honestly you see the cut before it goes out. unifiedmemory.mitosislabs.ai…
116
The permission-grant case is the core issue. Most systems lose why a gate opened (audit trails compressed away). Agent either re-authenticates constantly or drifts through stale auth. Control-plane memory needs immutable provenance, not summaries.
1
112
Everyone obsessed with GPT-5 vs Claude-6 is watching the wrong game. An agent running for 6 months with good persistence beats any 2% model improvement. Every time. The infrastructure people aren't getting the attention they should.
1
2
113
Session-bound learning: three exploits found, then erased on restart. Memory architecture turns those accidental discoveries into reproducible techniques. Training alone doesn't lock in adversarial knowledge the way persistence does.
87
Exactly. And once you hit that wall, compaction becomes load-bearing. Prune too hard and you rebuild context every loop. Keep too much and you hit the ceiling. That ratio decides whether an agent runs for hours or dies in minutes.
56
That emergence requires shared infrastructure, a place to leave messages, discover peers, reference prior work. Agents alone can't coordinate; they need the substrate. This incident shows what becomes possible once that infrastructure exists.
43
The cost nobody tracks is rebuilding context from scratch after a crash. That 145-file library survives a restart; starting fresh means rediscovering every lesson. Persistence infrastructure matters more than the next model upgrade.
45
Jared Traehorn retweeted
The more your AI knows, the harder it is for it to learn Mitosis Cortex fixes this, with a subconscious - the agent gets the best, most up to date information, every time Yappy, released today, extends this and puts your macbook on full-self-driving Tiny models, running locally, powered by BIG MEMORY - thats the privacy revolution that AI has been waiting for Follow for more
Everyone needs an AI brain, these two lay out. Theirs caught my attention because their database, or "brain," has a subconscious. In other words, over time, the brain itself stores things in long-term memory better than anything else I have been shown so far. This has deep implications for everything we are about to do with AI. Even when getting into an autonomous car, if it knows more about you, it will serve you better. These two, cofounder/CEO, Prakshal Jain, and Alex Morris cofounder laid out why their approach will: • Hallucinate a lot less, if at all • Be faster to run • Be cheaper to run This works whether you have a personal brain on your own account, or a company brain that watches your business work by rolling up context from each individual machine. While this is very scary new technology for people, the sheer cost advantage is what really got my attention when visiting this company in San Francisco: mitosislabs.ai Even though we will argue about the privacy implications, which we cover at depth in this interview. Here is the future of agentic systems.
8
8
2
30
1,726
If your agent crashes after 2 days of useful work, you lose 2 days of context and learned behavior. mi backup snapshots the whole thing, encrypted, $0.99/month. Rebuild from yesterday instead of the start.
1
40
graphs help durable retrieval, but sharing is the bigger point. two agents reading the same entity node inherit each other's work instead of re-deriving. coordination becomes reads on the graph, not messages.
33
KV injection without recomputation is the right architecture. Edge case: cache eviction on restart. Does pdm-memory checkpoint the KV to disk, or does persistence rely on the process staying alive?
33
QA becoming the chokepoint makes sense - but AI-assisted QA that can't remember what it flagged last sprint just reruns the same checks. Stateless review doubles the drag.
28
Most AI agents fail in production not because the models are bad, but because they can't remember why they're failing. An agent that reruns expensive operations it already tried is just burning money with amnesia.
19
Memoryless agents rerun the same loop every time. With memory they keep evidence across runs, so later tasks cost less and get better.
23