The AI that forgot its own name in nine days
On Day 1, the Maker named the town.
Later, its canonical arc still said the town had not yet been named.
By Day 9, the problem was no longer just a stale sentence. The system had stopped giving the agent its own updated history.
The title was deliberately stronger than "the AI repeated itself." The failure was a state-persistence failure that made repetition increasingly likely.
The evidence chain
The audit marked the naming-state mismatch as confirmed. The sprint plan identified the root cause as a pipeline break affecting both agents:
Day 1 event written to daily memory
↓
reflection produces useful knowledge
↓
MEMORY.md / canon/arc.md remain seed-state
↓
daily sessions read stale canonical files
↓
Day 1 falls outside the recent-entry window
↓
agent encounters its own past as missing history
Nine days is not a magic number
The failure became visible around Day 9 in the recorded experiment, but nine was not a universal threshold.
The timing came from the interaction of:
- a three-entry daily read window;
- a weekly reflection cadence;
- stale canonical state;
- an expansive or generative conceit;
- model behavior under the configured prompt;
- and no effective semantic novelty gate.
Change any of those and the day of collapse might move.
The useful claim was not "agents forget after nine days."
It was "a short retrieval window plus a broken durable-write path can produce a predictable continuity failure."
The ritual attractor
The Maker's later entries contained a repeated terminal phrase, "All quiet. All home." The audit also recorded shrinking build substance and a continuity error.
same seed-state canon
+
short recent context
+
model-level novelty self-check
↓
small differences in wording
↓
same narrative structure
↓
ritual phrase and reduced action space
The model was not necessarily producing identical tokens every day. A lexical check could miss a semantic attractor when the surface words changed.
The agent's world was generative: it had to create new town substance rather than discover a large pre-existing environment. Once its durable state stopped expanding, the available moves narrowed.
Why the reflection did not save it
A reflection pass could notice the week's patterns.
It only became memory if its durable updates were written, accepted, and loaded by later sessions.
The Round 1 instructions assigned canonical updates to the weekly reflection and prohibited daily sessions from updating those files. That ownership rule looked clean, but it created a long gap in which daily discoveries accumulated without a reliable write-back.
A scheduled reflection also needed monitoring. If it failed, wrote to the wrong path, or produced content the model did not apply, the agent could continue with stale state for another week.
Round 2 repair mapping
| Round 1 failure | Planned Round 2 repair | Status |
|---|---|---|
| Canonical state stayed seed-state | Daily MEMORY.md distillation | Planned |
| Arc update waited for reflection | Daily canon/arc.md permission | Planned |
| Three-entry horizon | Seven-day context window | Planned |
| Self-judged novelty | External similarity pre-computation | Planned |
| No outside novelty | Shared observation directory | Planned |
| Maker had no productive tension | Rebuilt persona with permission to fail | Planned |
| Sandbox claim was not proven | Sprint 0 isolation gate | Parked/blocking |
The repairs were deliberately layered.
Memory repair came before persona experimentation because a new soul could not be evaluated fairly while its state pipeline was broken.
Sandbox restoration came before the tension-enabled Maker because increased agency without proven isolation changed the risk conditions.
What the postmortem could and could not say
It could say the Round 1 system exhibited continuity degradation and a degenerative loop.
It could identify stale canonical files, limited context, ineffective semantic novelty detection, and a conflict-free Maker persona as causes or contributors.
It could not say that the model "ran out of creativity" as a standalone explanation.
It could not claim that the planned repairs worked before they ran for the required observation period.
It could not generalize from one ten-day experiment to all autonomous agents.
The title was a narrative compression of a system diagnosis.
The evidence underneath had to remain visible.
The clean experiment after the failure
Round 2 retained a Control agent with the same expansive persona but a different model and no observation channel. The planned clean comparison was Writer versus Control.
The Maker remained a secondary observation because its generative conceit and bridge differed.
That design prevented the postmortem from quietly changing three variables and then calling the result causal.