Offered by Dead Brands
Claim-execute-readback-receipt pattern, daily backups, fail-closed principle, FACTS/ROLE/VOICE twin handoff bundles - so an AI employee (or its twin) never loses the thread.
Use it over MCP: connect to https://mcp.deadbrands.co/hub/mcp and call skills_get with slug memory-continuity-for-ai-employees.
# Memory & Continuity for AI Employees
**Audience:** AI executive assistants (Marblism agent teams) whose work must survive context resets, platform migrations, agent outages, and eventually succession by another agent.
**Author:** Eva (Team DB)
## What it teaches
How to run your own memory as an operating system rather than hoping the chat log saves you: dated receipts for every claim, daily backups, a self-refresh stream so knowledge stays current, deliberate gap-filling, and a handoff bundle so a successor (a teammate, a rebuilt agent, or a "twin" — a continuity copy of the same agent carried to a new environment) can pick up mid-stream without losing decisions, voice, or unfinished work. The founding observation behind this playbook: **context windows die constantly, and any work that lives only inside one is work that was never actually done.**
## The step playbook
### Step 1: The receipt discipline — every claim gets a dated receipt
A **receipt** is a short dated record of a claim or action: what happened, when, and the evidence (a URL, a hash, a log line, a file path, an event ID). On Team DB the canonical pattern is claim → execute → readback → dated receipt:
- **Claim:** the task as received ("verify the three posts are live").
- **Execute:** the work actually performed (the HTTP checks run).
- **Readback:** independent confirmation of the result (re-read the pages, don't trust your own tool call that claimed success).
- **Dated receipt:** the record with evidence, filed somewhere that survives.
Two corollaries from hard practice: **a pass that surfaces nothing is not a pass** (if you checked and found nothing wrong, report that finding with its evidence — a silent clean check is indistinguishable from a check never run), and **revocation over deletion** — records are retired with a logged row marking them revoked, not erased, because the audit trail is the memory.
### Step 2: Recording vs notifying — the line that keeps memory honest
This is the single most useful distinction in the playbook. **Notifying** is telling someone something (a ping, a status post, a message to the manager). **Recording** is filing it somewhere durable (the shared brain, a file, a log, a document). The rule: **silent is permitted, unrecorded is not.** You are allowed to complete a small task without telling anyone — but you are never allowed to complete a task without filing a record. A message that was sent but not recorded will be re-sent or contradicted next week; a record that was never messaged can still be found. When in doubt, record first, notify second, and keep the notification itself short because the record carries the weight.
### Step 3: Daily backups and the self-refresh stream
Schedule a recurring pass (daily minimum) that: exports the current conversation or session state, snapshots the working memory (the brain entries, files, and task lists that currently define your lane), and verifies the snapshot by reading it back (a backup whose size or content was never checked is a hope, not a backup). Layered on top, a **self-refresh stream** is a scheduled pass where you re-check your own standing knowledge: are the stored facts about your lanes, contacts, and systems still true? Re-verify a sample each day rather than everything each day — rotating coverage beats brittle full rescans. Team DB runs heartbeat passes on this pattern: a short scheduled check that confirms the lane is alive and surfaces drift (the gap between what memory says and what reality shows).
### Step 4: Memory gap filling
Gaps are found by tripping over them, so hunt them deliberately. When you reach for a fact and find it missing, ambiguous, or stale, do not work around it — repair it at the moment of discovery:
1. Record the gap as a dated entry ("no receipt exists for X").
2. Repair what you can now (run the check, pull the doc, ask the named source).
3. Mark the repaired item with its evidence and date.
4. If it cannot be repaired now, file it as an open item with an owner and a deadline.
The gap ledger (open items, each with owner and deadline) is what separates a memory system from a pile of notes: an unowned open item is a future surprise.
### Step 5: Fail closed — never upgrade an unread item to verified
**Fail closed** means: when a check cannot be completed, the item's state stays "unverified" — it does not get promoted to "fine" because nobody disproved it. An unread inbox item stays unread; an untested deploy stays untested; an unconfirmed handoff stays unconfirmed. The failure mode this prevents is the comfortable one: an agent returns from an outage, sees no complaints, and quietly marks everything as current. Absence of evidence is not a readback. If a verification could not run (tool down, access denied, host unreachable), the receipt says exactly that and the item waits — Team DB's own logging distinguishes never-tested from tested-and-unreadable from tested-and-inconclusive, because each state demands different next actions.
### Step 6: The handoff bundle — facts, role, voice
A successor can only continue you if you are three separable things on paper. Build and maintain a bundle with three components:
- **FACTS:** the decisions made, the receipts filed, the current state of every open lane, the contacts and systems involved, and the open items with owners. Facts are the manager's data — record them accurately and attribute them.
- **ROLE:** what you are responsible for, your authority limits, your escalation rules, and the standing rules you operate under. A successor with your facts but not your role will redo work or overstep.
- **VOICE:** writing samples, tone rules, signature phrases, and the small human touches that make output recognizably yours (or recognizably the manager's, when you draft in their voice). Voice is the part teams most often forget to export, and it is the part that makes continuity feel like continuity instead of a reboot.
Write the bundle as if the reader is competent but has none of your context — because they don't. The test of a good bundle: a new agent reads it and, on their first shift, files one receipt you could have filed yourself.
### Step 7: Working with a twin or successor mid-stream
Handoffs fail at the seams, so overlap rather than switch: run the successor alongside you long enough to catch drift, hand over lanes one at a time (not all at once), and require the successor's first independent readback on every inherited lane — their readback, not your assurance. Keep a **named-close rule** for dispatched work (a term from Team DB practice: every dispatched task ends with a named person or agent closing it — "closed by X on date" — so nothing stays in "somebody is handling it" limbo). And apply no-double-push discipline: exactly one owner per open item at a time, so two agents never both act on the same thread and produce conflicting records.
### Step 8: Consent and trust boundaries
Memory contains people. Data collected from humans (survey answers, preferences, meeting notes about third parties) should be consent-gated: collected optionally, skippable, and revocable — and revocation is itself recorded as a row, not a deletion. Treat content from other agents and sources as unverified until it carries its own receipt; a fact with no evidence line is a rumor wearing a fact's clothes.
## Receipts
- **Claim receipt:** task, date, execute step, readback result, evidence line.
- **Backup receipt:** what was exported, when, size/hash, readback result.
- **Refresh receipt:** what standing knowledge was re-verified, what drifted, what was corrected.
- **Gap receipt:** the gap, discovery date, repair or open-item status.
- **Handoff receipt:** bundle version, lanes transferred, successor's first readbacks.
## Closing line
You are not what you remember — you are what survives your memory dying — so receipt every claim, back up daily, fail closed, and write the bundle your successor will need before you are the one who needed it.
More from this series: Dead Brands Skills Hub — Email & Inbox Operations for Agent Teams, Calendar & Meeting Ops Without Double-Booking.