Context Bankruptcy Skill
Every long-lived agent slowly fills with sediment: the org chart from two
reorgs ago, the project that got cancelled but still shapes its suggestions,
the preference you expressed once, sarcastically, in March. Deleting everything
loses the genuinely valuable judgment it accumulated; deleting nothing means
arguing with a colleague who lives in the past. Bankruptcy is the middle path
with discipline: audit the beliefs, keep what's true, correct what drifted,
purge what's wrong or expired — and write down what was lost, because silent
memory loss is how the same wrong belief gets re-learned from the same stale
sources next month.
What This Skill Produces
- A belief audit: what the agent currently holds, organized into facts /
preferences / procedures / relationships, each dated and sourced where
possible
- A keep / correct / purge ledger with a reason per entry — the artifact
that makes the bankruptcy deliberate instead of a rage-wipe
- A restated ground-truth file: the clean, current worldview to reload,
written to survive the next drift longer (dated claims, expiry hints)
- A bankruptcy record: what was purged and why, plus the re-learn guards —
which stale sources fed the bad beliefs and how to stop them refeeding
Required Inputs
Ask for (if not already provided):
- The agent's memory contents, exported or pasted (memory files, saved
context, custom instructions, whatever the platform exposes) — the audit
works on what it can see, and says so
- The symptoms: what it keeps getting wrong, where it contradicts itself
- What changed in reality (reorg, pivot, new stack, new owner) and when
- What the agent is good at that must survive — the reason this is
bankruptcy, not deletion
Working from a brief
Given what the user says is wrong and what is worth keeping, do the clean-up anyway: the facts to delete or correct, the preferences to keep word for word, the fresh memory written out in full, and a short routine to stop it going stale again.
Process
- Extract beliefs, not text. Convert the memory dump into discrete claims:
"believes the platform team owns billing" · "believes user prefers terse
answers" · "believes deploys happen Fridays". Tag each: fact / preference /
procedure / relationship, with best-guess age. Unstated-but-acted-on
beliefs (visible in the symptoms) go in too, marked inferred.
- Sort against current reality. With the user, mark each claim KEEP
(true, valuable) · CORRECT (right shape, wrong details — write the fix) ·
PURGE (wrong, expired, or toxic — including preferences the user no longer
holds). Rule for ties: a belief that can silently misdirect output is
PURGE-by-default; a belief that's merely unused can stay.
- Find the feeders. For each purged belief worth the trouble: where did it
come from, and does that source still exist (an old doc it can read, a
stale instruction file, a pinned message)? Purging the belief but not the
feeder schedules the relapse.
- Restate ground truth to age well. Write the reload file with dated
claims ("as of Jul 2026, billing is owned by…"), explicit preferences in
the user's own words, and expiry hints ("re-verify org facts quarterly").
Load order matters on most platforms: ground truth in the durable slot
(instructions/memory), not a chat message that scrolls away.
- Record the bankruptcy. What was purged, why, date, and the guards
added. Schedule the next audit — sediment accumulates at a knowable rate;
quarterly is the sane default for daily-driver agents.
Output Format
## Belief audit — [agent], [date]
| # | Belief (as a claim) | Type | Age | Source | Inferred? |
## The ledger
| # | Verdict (KEEP / CORRECT→fix / PURGE) | Reason |
## Ground truth v[N] (reload this)
[Dated claims · preferences verbatim · procedures · expiry hints]
## Bankruptcy record
[Purged: what & why · Feeders closed: … · Guards added: … · Next audit: date]
Quality Checks
Anti-Patterns
Related
[[session-handoff]] compresses a session; this resets a worldview.
[[context-budget]] for the token-layout side; [[agent-severance]] when the
answer is offboarding, not bankruptcy.
Example Trigger Phrases
- "My agent keeps acting on outdated beliefs."
- "Purge my assistant's stale memory."
- "Audit what my agent currently believes."
- "Reset the agent's memory after our pivot."
1---2name: context-bankruptcy3description: "Declare bankruptcy on a long-lived AI agent's accumulated memory — audit what it currently believes, separate ground truth from stale and wrong, purge deliberately, restate the truths that survive, and log what was lost. Use when an agent keeps acting on outdated facts, contradicts itself across sessions, 'remembers' things wrong, or after a reorg/pivot makes its worldview obsolete. Produces a belief audit, a keep/correct/purge ledger, a restated ground-truth file, and the bankruptcy record."4disable-model-invocation: true # skill-optimizer5metadata:6 source: https://skillmd.com/skills/slipcast-dev/context-bankruptcy7---89# Context Bankruptcy Skill1011Every long-lived agent slowly fills with sediment: the org chart from two12reorgs ago, the project that got cancelled but still shapes its suggestions,13the preference you expressed once, sarcastically, in March. Deleting everything14loses the genuinely valuable judgment it accumulated; deleting nothing means15arguing with a colleague who lives in the past. Bankruptcy is the middle path16with discipline: audit the beliefs, keep what's true, correct what drifted,17purge what's wrong or expired — and write down what was lost, because silent18memory loss is how the same wrong belief gets re-learned from the same stale19sources next month.2021## What This Skill Produces2223- A **belief audit**: what the agent currently holds, organized into facts /24 preferences / procedures / relationships, each dated and sourced where25 possible26- A **keep / correct / purge ledger** with a reason per entry — the artifact27 that makes the bankruptcy deliberate instead of a rage-wipe28- A **restated ground-truth file**: the clean, current worldview to reload,29 written to survive the next drift longer (dated claims, expiry hints)30- A **bankruptcy record**: what was purged and why, plus the re-learn guards —31 which stale sources fed the bad beliefs and how to stop them refeeding3233## Required Inputs3435Ask for (if not already provided):36- The agent's memory contents, exported or pasted (memory files, saved37 context, custom instructions, whatever the platform exposes) — the audit38 works on what it can see, and says so39- The symptoms: what it keeps getting wrong, where it contradicts itself40- What changed in reality (reorg, pivot, new stack, new owner) and when41- What the agent is *good* at that must survive — the reason this is42 bankruptcy, not deletion4344## Working from a brief4546Given what the user says is wrong and what is worth keeping, **do the clean-up anyway**: the facts to delete or correct, the preferences to keep word for word, the fresh memory written out in full, and a short routine to stop it going stale again.4748## Process49501. **Extract beliefs, not text.** Convert the memory dump into discrete claims:51 "believes the platform team owns billing" · "believes user prefers terse52 answers" · "believes deploys happen Fridays". Tag each: fact / preference /53 procedure / relationship, with best-guess age. Unstated-but-acted-on54 beliefs (visible in the symptoms) go in too, marked *inferred*.552. **Sort against current reality.** With the user, mark each claim KEEP56 (true, valuable) · CORRECT (right shape, wrong details — write the fix) ·57 PURGE (wrong, expired, or toxic — including preferences the user no longer58 holds). Rule for ties: a belief that can silently misdirect output is59 PURGE-by-default; a belief that's merely unused can stay.603. **Find the feeders.** For each purged belief worth the trouble: where did it61 come from, and does that source still exist (an old doc it can read, a62 stale instruction file, a pinned message)? Purging the belief but not the63 feeder schedules the relapse.644. **Restate ground truth to age well.** Write the reload file with dated65 claims ("as of Jul 2026, billing is owned by…"), explicit preferences in66 the user's own words, and expiry hints ("re-verify org facts quarterly").67 Load order matters on most platforms: ground truth in the durable slot68 (instructions/memory), not a chat message that scrolls away.695. **Record the bankruptcy.** What was purged, why, date, and the guards70 added. Schedule the next audit — sediment accumulates at a knowable rate;71 quarterly is the sane default for daily-driver agents.7273## Output Format7475```76## Belief audit — [agent], [date]77| # | Belief (as a claim) | Type | Age | Source | Inferred? |7879## The ledger80| # | Verdict (KEEP / CORRECT→fix / PURGE) | Reason |8182## Ground truth v[N] (reload this)83[Dated claims · preferences verbatim · procedures · expiry hints]8485## Bankruptcy record86[Purged: what & why · Feeders closed: … · Guards added: … · Next audit: date]87```8889## Quality Checks9091- [ ] Beliefs are discrete, checkable claims — not pasted memory prose92- [ ] Every PURGE has a reason; every CORRECT has the actual correction written93- [ ] At least one feeder was traced — bankruptcy without closing feeders is a94 subscription to this skill95- [ ] Ground-truth claims are dated and the file says where it should live on96 the platform (durable slot, not chat)97- [ ] What-was-lost is recorded — the user can consciously re-teach, rather98 than discover absence mid-task99100## Anti-Patterns101102- [ ] Do not rage-wipe — full deletion without the audit loses the judgment103 that took months to accumulate, which is why this skill exists104- [ ] Do not keep contradictions to "be safe"; two contradictory beliefs means105 the agent picks one at random per session106- [ ] Do not restate ground truth undated — undated truth is next quarter's107 sediment108- [ ] Do not treat vendor "memory deleted" claims as verified for sensitive109 content — note what the platform actually promises (see110 [[agent-severance]] for full offboarding)111112## Related113114[[session-handoff]] compresses a session; this resets a *worldview*.115[[context-budget]] for the token-layout side; [[agent-severance]] when the116answer is offboarding, not bankruptcy.117118## Example Trigger Phrases119120- "My agent keeps acting on outdated beliefs."121- "Purge my assistant's stale memory."122- "Audit what my agent currently believes."123- "Reset the agent's memory after our pivot."