Karpathy+

00TL;DR

The goal: a shared memory between you and your AI that survives every session, updates when things change, and doesn't quietly fail.

Three layers (sources, wiki, schema). Three operations (ingest, query, lint). One hard boundary (knowledge vs runtime). Enforcement by defense in depth: behavior rules, a session hook, scheduled lint, git history, and a constitutional completion gate. Progressive disclosure keeps the per-session cost low: schema plus index plus two to four pages, not the whole corpus.

01Why Most Systems Fail

LLM memory systems tend to die in the same three ways. Every design decision here is a response to one of them.

02Three Layers

Knowledge flows upward with increasing structure and decreasing volume.

• Sources (wiki/sources/) are raw inputs: clipped articles, meeting transcripts, tool-intel feeds, bookmarks. Messy. The LLM synthesizes upward.
• Wiki (wiki/) is the curated layer. Target: 15 to 30 pages. LLM-maintained.
• Schema (CLAUDE.md, ~50 lines) is behavior rules plus wiki conventions. No knowledge lives here.

03Boundary & Buckets

One rule prevents the three-copy drift problem before it starts:

Knowledge goes in ~/AI/. Runtime goes in ~/.claude/.

Inside ~/AI/, no loose files at root. Every file lives in a bucket: wiki/, project repos, career/, personal/, OUTPUTS/, CLIENT-OUTPUTS/, inbox/ (untriaged), scratch/ (throwaway). Archive and backups live at ~/Archive/, outside ~/AI/ entirely, because the wiki's MCP is scoped to ~/AI/ and archive inside would pollute every search.

04Three Operations

Ingest

Triggered after deploys, debugging, corrections, decisions. The LLM updates relevant wiki pages and appends to log.md with a Judgment: line explaining the reasoning behind a fix or tradeoff.

Judgment: lines are not decoration. They are the retrieval path that makes the system self-correcting. A future Claude reading the log doesn't just see what happened, it sees what constraints are still binding. Load-bearing reasoning in Judgment: lines means the system resists drift from the inside.

Query

Triggered at the start of every session. The LLM reads INDEX.md and picks two to four relevant pages. Total injected context: about 50 lines of schema, 30 lines of index, the small set of relevant pages. The old failure mode (loading four hundred lines of rules before the human types) becomes structurally impossible.

Caveat for plan execution: when the task is executing a plan, the plan is not self-sufficient context. Load the project pages the plan references before taking its "next steps" as scope. Plans decay faster than the project state they depend on.

Lint

Runs weekly or on demand. A single slash command performs queue processing, freshness check, broken-link scan, task aggregation, log-gap check, source scan, log rotation, auto-memory drift check, runtime audit, and heartbeat stamp. Output: auto-fix pass plus triage report. Human responds do it / defer it / kill it.

Lint has a second life as a pattern detector. When it surfaces the same issue twice, that is the signal to build a hook or a guard. Rule: flag once via lint, build infrastructure only when the pattern fires twice. Prevents over-engineering on one-time problems.

05Defense in Depth

Discipline alone fails during fires. Multiple independent safety nets, each cheap to run always.

Output verification is the enforcement mode that closes the runtime-state gap. Lint reads files and cannot see whether a cron or launchd agent is actually loaded and running. An output verifier checks "did the scheduled job produce today's file, on time?" Any scheduled agent should declare an expected output path.

06Propagation is a Snapshot

A small principle with wide reach: copies detach from their source at propagation time. Verification is required at consumption, not at creation.

• Auto-memory (~/.claude/projects/*/memory/) is a copy. If it and the wiki disagree, the wiki wins. Lint step diffs key facts between them.
• UI-surface memories (Cowork global instructions, IDE settings) are copies pasted at a point in time. Schema changes require active propagation to each surface.
• Forwarded session instructions are copies of a prior session's verification state. Treat a forwarded instruction as a claim, not a direction. Re-derive scope from current wiki before executing.

Where possible, remove duplicated facts from copies and point them to the wiki. A hardcoded page count will drift. A link cannot.

07Hard Limits

08Build It Yourself

The core pattern is general. Anyone can adapt it.

Minimum viable wiki, five steps

1. Create wiki/ with INDEX.md and log.md.
2. Write a lean CLAUDE.md (~50 lines): style rules, "read INDEX.md first," update conventions, the boundary rule.
3. Create three to five pages covering your most-repeated context.
4. Set up lint. Even a manual weekly review counts at the start.
5. Use it for two weeks before adding complexity.

Keep these (universal)

Three layers, boundary rule, INDEX.md as entry point, log.md with Judgment: lines, lint as scheduled enforcement, page template for stable parse, defense in depth, completion gate.

Adapt these (local)

Page categories to your domains, lint schedule and notification medium, task tracker of choice, surface routing (skip entirely if single-surface), domain isolation (only matters with multiple clients).

The system grows naturally. Sources accumulate. Pages get added when you notice yourself repeating context. Lint catches what you forget. Everything beyond the minimum is earned by need, not added by default.

09What's Next

Planned, not built.

• last_reconciled: frontmatter for plans. A plan carries both last_updated (any edit) and last_reconciled (references re-verified against current project state). A lint step flags any plan whose last_reconciled predates the last_updated of a project page it cites. Cost: one frontmatter key plus one lint step. Deferred until the next plan-execution session consumes it.
• Richer auto-memory audit. Current drift check compares page counts, paths, project names. A full audit would diff every factual claim in auto-memory against the wiki.
• Local search over the wiki (qmd or equivalent) once page count crosses ~50. Grep holds below that.

Everything else is earned by need, not added by default. The system is small on purpose.

The system is working well, and building it has been a genuinely valuable learning experience. That said — I'm actively looking forward to replacing most of it.

10Inspirations

The compounding principle (Atlas Forge) is the spine. The wiki pattern (Karpathy) is the skeleton. The boundary rule, defense in depth, heartbeat, and completion gate are the load-bearing response to failure modes each source acknowledged but did not fully solve.