We present the Alive Context System (ALIVE) — the first implementation of Personal Context Management. Built over eight months from plain Markdown files and bash hooks, ALIVE has grown from a Google Drive folder structure to a 9,754-line agent operating system with over 100 users in its first two weeks of public release. This paper documents the full evolution: five major versions, 567 commits across 23 repositories, 97 development sessions, and the architectural decisions that shaped each transition — traced to their origins in voice recordings, whiteboard sessions, and live user feedback. We present ALIVE not as a finished product but as living proof that the PCM category is viable, the Context as Code pattern works, and context ownership is achievable with technology that already exists.
In 2023, Ben Flint was in the top 1% of ChatGPT users worldwide — 5,407 conversations in a single year. The frustration was specific and persistent:
"I was in the top 1% of ChatGPT users last year, and I still wasn't able to get it to write me an email."
The problem wasn't intelligence. The model was capable. The problem was context. Every conversation started from zero. Every session required re-explanation. The model had no memory of decisions made, relationships built, or work completed. It was, in Flint's words, "an NPC stranger."
"The thing that makes AI stupid is it doesn't know your clients. It doesn't know what happened today."
Flint had used Tiago Forte's PARA framework (Projects, Areas, Resources, Archives) for personal knowledge management. It worked for organizing files. It didn't work for provisioning context to AI agents.
"PARA was designed for personal knowledge management. ALIVE was designed for operators running multiple businesses."
The frameworks available — PARA, Zettelkasten, GTD — all answered the same question: where do I file this? None of them answered: what does my agent need to know right now?
On January 3, 2026, standing at a desk in the middle of the night, Flint rearranged the letters of his framework and they spelled a word:
"I rearranged the letters and it spelt alive. And I was standing over at a desk in the middle of the night, and I was just like, What the fuck."
Archive. Life. Inputs. Ventures. Experiments. Five domains. One acronym. Each letter is a verb — you archive, you live, you intake, you venture, you experiment. The file system is the methodology.
On the same day, the paradigm got its name:
"I'm kind of calling this context as code, because we are running a software system without the software."
The genesis. A Google Drive folder structure with a CLAUDE.md rules file. Nine skills prefixed /z (for "zero"). No hooks. No voice specification. No persona. Five principles that would survive every subsequent version:
The install path was systemzero.com/operators. The tagline: "Context infrastructure that gives AI perfect memory of your work."
32 commits in the system-zero repo. The first commit message: "Initial commit: System Zero agents monorepo."
The rebrand. System Zero became ALIVE. The _state/ folder was renamed to _brain/. Hooks were introduced for the first time — two kept (session-start.sh, activity-log.sh), two cut (pre-compact.sh because "can't do agentic work," post-tool-use-files.sh because "nothing consumes the data").
12 skills planned. The skill prefix changed from /z to /alive:. The first hooks were built "in plan mode, which is not the right way to do it" — discovered from examining other plugins.
70 commits in the aliveskills repo across 11 branches.
The heavy version. Named in a session with Attila Mora:
"Elephants never forget. And they're long, stable, their pack animals."
"Is it unlimited elephant? Limitless elephant."
"It also means you're breaking the chains."
The voice specification appeared: "70% Sage, 30% Rebel. Qui-Gon Jinn spec." The kill list grew — "Unlimited Elephant" itself would later be killed when the system shed weight.
The first external user arrived. Attila Mora — creative collaborator, the first person outside the core team to run ALIVE — used it for 48 hours straight: 56 Claude Code sessions, 142MB of transcripts, $4,200 in API credits. He connected 9 MCP servers in one sitting. He built the ALIVE Daemon — a 770-line Python Telegram bot.
And he named the agent:
"The agents are squirrels. We're literally building wealth."
The squirrel metaphor — agents that scatter-hoard context like squirrels gathering walnuts — became the system's identity. The walnut became the unit of context. The metaphor was never planned. It emerged from a late-night lock-in session on February 18, 2026.
The public release. 14 hooks across 5 lifecycle events. 15 skills. 6 rule files. 114 files shipped in a single commit. The system folder was .walnut/. The install command: claude plugin install walnut@stackwalnuts.
The architecture:
_core/ kernel with 5 files: key.md, now.md, tasks.md, insights.md, log.md_capsules/ for work units with companion.md manifestssquirrel.core@1.0 runtimeWill Bainbridge — the co-architect who had been on weekly calls since June 2025 and participated in the defining whiteboard session on July 22, 2025 — reviewed every branch. His observation that "95% of sessions end without firing the hook" changed the save architecture.
Shipped one day after v1. The architecture overhaul:
_core/ became _kernel/ — three source files (key.md, log.md, insights.md) plus one generated projection (now.json)companion.md became context.manifest.yaml — structured YAML instead of freeform markdownPeople/The naming shift: .walnut/ became .alive/. The install command: claude plugin install alive@walnut.
9,754 lines of configuration across 57 files. Zero trained weights. An agent operating system built from:
| Component | Files | Lines |
|---|---|---|
| Rule files | 6 | 1,715 |
| CLAUDE.md | 1 | 127 |
| Skill files | 15 | 5,242 |
| Hook scripts | 15 | 1,835 |
| Template files | 19 | 835 |
Token cost per session: approximately 9,000 tokens — 0.9% of a one-million-token context window.
14 hooks across 5 lifecycle events create a persistent runtime from a stateless model:
Six files defining the behavioral constitution:
The system provisions context at three levels:
"There's two points where you orient the agent. The first is at runtime. You orient it to your world in a lightweight way. And then the prompt is the second orientation — let's load the context from a walnut and check what needs to be done."
And the bundle level:
"The capsule then becomes the rolling sum total of the external context gathered."
| Level | What loads | When |
|---|---|---|
| Runtime | World identity, rules, preferences (~9K tokens) | Every session start |
| Walnut | Kernel files, bundle manifests, tasks | On load-context invocation |
| Bundle | Manifest, observations, raw sources, bundle skills | When work begins on specific deliverable |
Two tiers of behavioral control:
The split: rules that must be absolutely enforced use hook guards. Rules that benefit from judgment use context injection.
Fifty-one items were explicitly killed across five versions. Each kill tells a story about what was tried, what failed, and what the system learned:
key.md._state/ → _brain/ → now.md + log.md at root → _kernel/ — the folder name journey. Each rename reflected a philosophical shift in what the system files represent.pre-compact.sh — killed in v2 design because "can't do agentic work." Later resurrected in v2.0 when the hook architecture could support it.56 sessions. 142MB. $4,200 in API credits. February 7-9, 2026.
Named the squirrel. Named the walnut. Co-authored manifesto sections. Built the ALIVE Daemon. Created Modes (Creative + Savant). Declared "No ads on alive. Ever."
"What I admire most about alive, it's not necessarily the file protocol system — stunning that, but I think a lot of engineers could come up with that. What I really admire is the philosophy of the way it flows from experiments, ventures, life."
2,226 mentions across 240 transcripts. Weekly calls from June 2025. The whiteboard session on July 22, 2025 — the most productive session in the project's history — established the file-based architecture.
Restored _core/ when Flint had flattened it. Authored the "Ambient Context" branch (16 changes cherry-picked to main). Observed that "95% of sessions end without firing the hook" — which changed everything about how saves work.
Built ALIVE for Hermes the same day he saw it — without help. Validated the squirrel visual language, the skill-based integration, and the walnut model. His tweet hit 500,000 views, naming ALIVE as "the most important part" of his Hermes Agent setup. Exposed untested v2 features, forcing a rollback to stable v1.
Two weeks since public release. No marketplace listings. No paid promotion.
| Metric | Value |
|---|---|
| GitHub stars | 39 |
| Unique cloners (14 days) | 257 |
| Unique visitors (14 days) | 241 |
| Peak day clones | 310 (March 27) |
| Known users | 100+ |
| Confirmed live users (March 27 standup) | 50+ |
| Viral tweet views | 500,000 |
| Press coverage | Zero |
| Marketplace listings | Zero |
Top referrers: GitHub (93 visits), walnut.world (90), Reddit (75), Threads (59), alivecomputer.com (50). International reach: approximately 15 of 39 stargazers are Korean developers, discovered organically.
Testimonials from early users:
"most cracked thing I've seen for AI in 2025." — Louka Ewington-Pitsos, AI Researcher
"everyone who uses this is gonna run circles around anyone else." — Athon Millane, Frontier Model Lab
"best thing ive ever used. this is fucked." — Caspar Tremlett, Brand Strategist
"I've literally been like hours sleep a night the past three nights because I'm just hooked on it." — Brenton Chan, organic user
Nine things from the very first version — System Zero, December 2025 — survived every rewrite, every rename, every architectural pivot, and exist in ALIVE v2.0 today:
| read indicators in every sessionEverything else changed. The names changed. The folder structure changed five times. The skill count went from 9 to 15. The hook count went from 0 to 14. The voice specification went from nonexistent to a configurable per-walnut system with energy matching and a sycophancy guardrail.
The things that survived are the principles. Everything else is implementation.
From a voice recording, early 2026:
"Context infrastructure is the only unsolved problem in AI in 2026."
And from the same recording:
"You are now a few markdown files away from outperforming every other person on the planet."