If your AI starts every session as a stranger, you already know the problem.
The people solving it now have a room.
ALIVE Discord: open, public, quiet enough to actually talk.
👇 https://t.co/aQCYwd8pTo
ALIVE v3.2.0 shipped three features that solve the same problem: personal context systems are hard to try, hard to migrate, and hard to run concurrently.
/1/ Demo Worlds
ALIVE stores your life: ventures, people, decisions, domain knowledge. Showing it to someone means exposing everything.
/alive:demo scaffolds a complete synthetic world in ~10 seconds. People, ventures, 6 months of log history, structural validation. Zero LLM calls on the deterministic path.
Custom path: describe a persona in one paragraph. A 5-stage pipeline builds a world around it.
Full lifecycle commands: activate, deactivate, delete, resume.
Try the full system. Evaluate it. Break it. Your real world stays untouched.
/2/ System Upgrade
File-based systems accumulate structural debt across versions. Stale directories survive. Retired patterns hide. Migration is the scariest command in any context tool.
/alive:system-upgrade is now a 13-phase orchestrator.
Version detection reads content fingerprints: three signals, lowest wins. Doesn't trust metadata. --dry-run previews every operation before it touches a file. --resume picks up from the last clean phase on failure. --rollback reverts.
Five retired patterns now caught that survived previous upgrades silently.
/3/ ALIVE CLI
bin/alive is a deterministic CLI layer. Log writes, doctor checks, task promotion, all through one interface.
Shared locking prevents two sessions from writing the same file. Error attribution traces every failure to the skill that caused it.
Foundation for concurrent sessions.
https://t.co/t5F83zhXYX
Three teams. Same week. Same conclusion.
~ Anthropic shipped Dreaming. Background process that reviews past agent sessions, extracts patterns, writes plain-text memory entries. Harvey reported 6x task completion in pilot.
~ Mem0 published State of AI Agent Memory 2026. Selective memory at 7k tokens beats full context at 26k+. Less when structured.
~ Letta red-teamed the Context Constitution. Models assume ephemerality at a level prompting can't override. External memory writes are the architecture.
while ALIVE ships this loop as cron jobs and plain files. Background processes write to walnuts. Sessions read from walnuts. Next session compounds.
Open source.
https://t.co/e5lI7GTYmF
ALIVE v3.2.0 🐿️
🏗️ /alive:demo - scaffold a full world from one paragraph
🔧 /alive:system-upgrade rewritten - 13 phases, --dry-run, --resume, --rollback
🔍 version detection via content fingerprinting
⚡ alive CLI for log writes, doctor checks, task promotion
🛟 failure recovery - partial runs resume from last good stage
66 commits. 79 files. Plain files, still yours.
https://t.co/t5F83zhXYX
Your AI forgets you every conversation.
The intelligence is real. The amnesia is structural.
The platforms "fixed" this by building memory products. OpenAI's memory. Anthropic's Projects. Gemini's context.
Each one a garden with a wall. Your memory in their system, queryable through their interface, portable nowhere.
That is the original sin of the platform era, applied to AI.
~~~~
We propose a different starting point. Context is property.
Your emails. Your decisions. Your relationships. Your projects. Your history. They are the substance of your working life.
Context stored as plain files on your machine has properties no platform memory has:
> Permanent: Files do not expire when companies pivot.
> Portable. Any AI that reads files can use it.
> Auditable. Open it in any text editor.
> Composable. Link, fork, extend, merge.
Five domains: Archive, Life, Inputs, Ventures, Experiments. The unit is a walnut.
A folder with three files.
- key.md holds identity.
- now.md holds the current state.
- log.md holds the history, prepend-only.
A new session opens a walnut and arrives with context. It does not need to be briefed. It already knows. At the end, decisions log, tasks route, state updates. The next session reads what this one wrote.
The AI that works with you on day ninety knows more than the AI on day one. Not because the model improved.
Because the context did.
Everyone built the AI wrapper. We built the human wrapper.
Not a product. Infrastructure.
Open source.
Plain files.
Yours.
https://t.co/6Su26qqHSX
tried a lot of memory tools. ALIVE is the only one where the context travels properly between setups.
it creates 5 folders on disk. one walnut per subfolder. everything else is markdown.
for instance: it scanned my Hermes config on the Mac Mini and dropped md files straight into my Claude cowork folders. 30-second handoff.
ask me for my Hermes setup now and you get one capsule. readable on any agent.
Claude Projects can't migrate between Opus and Sonnet. walnuts don't care which model is reading them.
I could also send my content walnut to someone helping with content: voice rules, style notes, past corrections... they'd write on-brand from paragraph one.
git pull and Oz's context is on a different machine. no sync service, no export wall.
Claude Projects stay in Claude. walnuts go wherever I go.
Try to export your ChatGPT memory.
Try to bring your Claude Projects to Gemini.
Try to run any of it on a new agent.
"Memory" they sell you is memory you don't own.
We fix this.
https://t.co/N1aSg1Vgyo
was on free tiers of Notion, Calendly, Otter, Cursor. the kind of setup you outgrow in a month.
2 months on ALIVE and nothing to outgrow. the limits don't apply when the context lives on your disk.
thread coming this week.