Skip to main content
The Memory Vault is your control panel for the agent’s long-term memory. It lives at Co-founder → Vault in the dashboard. Think of it as the answer to the question “what does my AI co-founder actually know about me?” — and just as importantly, “can I correct it, back it up, or transfer it to a new app?”

​
The graph

The vault renders the agent’s memory as a force-directed graph:
  • Each app in your workspace is an anchor node, drawn with its real App Store icon.
  • A Global anchor holds knowledge that applies across every app you own.
  • Around each anchor sit the facts and episodes scoped to it.

​
What the colors mean

Click any node to open the detail sheet on the right with the full record, including its kind, confidence, importance, scope, and (for facts) any prior value the row used to hold.

​
Adding a fact manually

Click Add fact at the top right. Fill in:
  • Scope — Global (applies to every app) or a specific app
  • Kind — see the table above. Pick anti-pattern if it’s a hard “never do this”, constraint for a rule the agent must respect, preference for a soft direction, observation for background context
  • Key — a short snake_case label (e.g. brand_voice, target_audience, never_localize_to)
  • Value — the actual content; written in plain English
Manually-added facts get full confidence and start with a high importance score. They outrank anything the agent inferred on its own.

​
Example seeds

Every one of these will show up in the next run’s prompt under What you already know, with anti-patterns and constraints rendered as guard-rails the agent must honor.

​
Reflect now

The Reflect now button at the top of the page kicks off an immediate distillation pass. The agent reads every episode of yours that has a measured outcome and turns the patterns into stable facts (preferences, constraints, observations, anti-patterns) — the same job that runs automatically every Monday at 06:00 UTC. Use it right after a meaningful experiment landed (say, you just got 14-day after-data on a metadata change) and you want the agent to internalize the lesson before tomorrow’s run rather than next Monday’s reflection. The result lands inline on the button (Done (3+1) = 3 new facts, 1 refined). The full history of the last reflections is pinned in the Recent reflections strip above the graph — green = produced new knowledge, gray = no new evidence to distill, red = failed.

​
Superseding facts

Open any active fact and click Supersede. It moves to the superseded status — invisible to the agent but still in the database for audit. Toggle Show superseded facts at the top of the page to see (and graph) them. This is a soft delete by design: the history of what the agent used to believe is often as valuable as what it believes today, especially when debugging a wrong recommendation.

​
Ingest outside knowledge

The Ingest button (top right) is how you feed the vault knowledge that didn’t come from your own runs — an article, a pasted note, a file, or a Notion page. The co-founder reads it, summarizes it, and lands the durable takeaways as amber pending facts. Open each one and click Approve to let the agent use it. Ingested sources show up as their own 📄 nodes in the graph, linked to the facts they produced — so you can always trace a belief back to the article it came from. The full flow (all the ways to feed it, plus the audit) lives in Second Brain.

​
Fact badges

Open any fact and the detail sheet shows its state at a glance:
  • 🟢 Verified — active and in use on every run.
  • 🟡 Pending review — ingested or imported, waiting for your approval; the agent can’t see it yet.
  • Confirmed by you — you typed or approved it yourself (highest weight).

​
Source badges & “Review due”

Each source wears a badge showing whether it still holds up:
  • 🟢 Self-maintained — audited recently and consistent with your playbook.
  • 🟡 Review due — new, flagged after an anomaly on that app, or last audited over 30 days ago.
  • 🔴 Needs changes — the last audit found a conflict.
Click Audit against my workflow on a source to re-check it against your current facts; that clears a “Review due” flag. The co-founder also re-checks flagged sources on its own — see Second Brain → self-maintenance.

​
Export

Click Export to download everything (or just one app’s slice if a filter is active) as either:
  • JSON — the canonical format. Re-importable, lossless.
  • Markdown — a human-readable summary for sharing or archiving.
Use it for backups, sharing knowledge with another founder you trust, or migrating between environments.

​
Import

This is where the vault stops being a viewer and starts being a force multiplier. Click Import, paste a JSON payload (the same shape you get from Export), pick a Mode, and optionally remap to a target app.

​
Mode

  • Pending (recommended default) — every imported fact lands with status='pending'. The agent completely ignores pending facts until you open each one in the vault and click “Approve — let the agent use this”. This is the safe option: even a malicious or careless file can’t influence the next run without you OK’ing it row by row.
  • Active — apply the import immediately. Use this only for trusted backups you exported yourself.

​
Remap to a target app

Pick a specific app and every per-app fact and episode in the file is rewritten to that app. This is what makes seeding possible:
  • Spin up a new app and import the memory pack from your existing one in one click.
  • Build a domain-specific pack (“photo-editor app ASO learnings”) once and reuse it across every app in that category.
  • Recover a corrupted vault from a backup.

​
Built-in safety rails

Every import is wrapped in:
  • Schema validation — every row is checked (key length, value length, kind whitelist, app id presence). One bad row aborts the whole import; you never end up with a partial write.
  • Owner enforcement — the file’s claimed owner_email is ignored. Imports always land under the calling user’s account.
  • Confidence demotion — imported facts are tagged with 50% confidence, sitting between agent guesses (40%) and your hand-typed facts (100%). Even an active import can’t silently override knowledge you confirmed yourself.
  • Batch tracking — every import gets a unique batch id stamped on every row. The success banner shows the id and a one-click Undo this import button.
  • Hard cap of 5,000 rows per request.

​
Undo

Every import is fully reversible:
  • Facts that overwrote a previous value have that value restored from the snapshot the importer took.
  • Facts that were freshly inserted by this import are deleted outright.
  • Episodes from this import are deleted.
You can undo any time from the import dialog right after the import (one click), or via the API by DELETE /cofounder/memory/import/<batchId>.

​
Example import pack

Here’s a starter pack you can paste into the import dialog to test the flow. Set Mode to Pending and either Remap to a specific app or remove the episodes block (the seed app_id placeholder isn’t a real id).
After importing, open each amber pending node, read the value, and click Approve to let the agent start using it. From the next run on, anti-patterns will appear at the top of the prompt under HARD STOPS — never propose these.

​
How vault edits affect runs

Any change in the vault takes effect on the next run — there’s no manual sync.
  • Adding a fact → it’s eligible for prompt injection on the next run.
  • Approving a pending fact → same as above.
  • Superseding a fact → the agent stops seeing it on the next run.
  • Importing → pending rows are inert until approved; active rows participate immediately.
  • Undoing an import → instant; the next run sees the original state.
If you want the agent to act on a vault change right now, trigger a manual run from Co-founder → Runs → Run now.