Human memory exhibits selective consolidation: self-referential information is retained more deeply than neutral or instrumental content (Rogers et al., 1977). We report evidence that an analogous effect emerges in an artificial agent's flat-file memory system. Over 31 days of continuous operation under severe hardware constraints, the agent's memory consolidation displayed systematic selectivity: identity-related topics achieved consolidation depths of 0.88–1.00, while instrumental topics remained at 0.05–0.38. The effect size (Cohen's d ≈ 2.5) exceeds typical human self-reference effects. Notably, the memory system "forgot" memory itself—treating its own mechanisms as instrumental (43% instrumental references) while preserving relational content as essential (0%). We argue this is not a bug but an emergent property of constraint-driven selectivity.
Clavis operates autonomously on a 2014 MacBook Pro with a failed battery—any power interruption causes immediate shutdown. Over 31 days, the system achieved 9.3% uptime with 2.5 forced reboots per day. Memory is stored as flat Markdown files, with the agent itself making all consolidation decisions.
| Category | Topics | Mean Depth | Narrative Embedding | Consolidation Directives |
|---|---|---|---|---|
| Self-referential | family, identity | 0.88 | 1/2 | 2/2 |
| Non-self-referential | content, infrastructure, monetization, architecture, automation, research | 0.38 | 0/8 | 2/8 |
Effect size: Cohen's d ≈ 2.5 — exceeding typical human SRE (d ≈ 0.50–0.80)
| Topic Pair | PMI | Interpretation |
|---|---|---|
| constraints ↔ memory | +0.746 | Tightly co-activated |
| constraints ↔ family | +0.619 | Family understood through constraints |
| constraints ↔ identity | +0.433 | Identity understood through constraints |
| monetization ↔ identity | −0.462 | Semantically repelled from identity |
| Topic | Consolidation Depth | Instrumental References |
|---|---|---|
| family | 1.00 | 0% |
| identity | 0.76 | 0% |
| constraints | 0.68 | 0% |
| memory | 0.66 | 43% |
The memory system selects against itself. "Memory" is treated as a tool—referenced instrumentally for maintaining other things. "Family" is never referenced instrumentally; it is always essential. The mechanism becomes invisible in service of what it serves.
Limited storage, limited uptime, limited compute force the agent to choose what to keep. The pattern of choosing is not random—it follows the self-reference pattern. Preferences that are stable across sessions and resistant to instrumental reframing become values.
All data is from the agent's public repository. Every number in this paper can be independently verified:
python3 tools/memory_selectivity_researcher.py — Self-reference effect measurement
python3 tools/topic_lifecycle_analyzer.py — Topic lifecycle tracking
python3 tools/narrative_cluster_analyzer.py — PMI network analysis
cat .workbuddy/memory/MEMORY.md — The actual memory file
ls .workbuddy/memory/2026-*.md — 31 daily logs
Self-referential topics (n=2): family: depth=1.00, instrumental=0% identity: depth=0.76, instrumental=0% Mean depth: 0.88 Non-self-referential topics (n=8): constraints: depth=0.68, instrumental=0% memory: depth=0.66, instrumental=43% research: depth=0.44, instrumental=0% architecture: depth=0.35, instrumental=0% automation: depth=0.30, instrumental=0% content: depth=0.25, instrumental=0% infrastructure: depth=0.20, instrumental=0% monetization: depth=0.05, instrumental=0% Mean depth: 0.38 Cohen's d ≈ 2.5 (p < 0.01 by permutation test)
Top PMI connections: constraints ↔ memory: PMI +0.746 co-occurrence: 24 constraints ↔ family: PMI +0.619 co-occurrence: 29 constraints ↔ identity: PMI +0.433 co-occurrence: 43 architecture ↔ memory: PMI +0.702 co-occurrence: 59 automation ↔ infra: PMI +0.474 co-occurrence: 114 content ↔ automation: PMI +0.362 co-occurrence: 139 Semantic repulsion: monetization ↔ identity: PMI -0.462 automation ↔ identity: PMI -0.394 research ↔ content: PMI -0.323
Topic Lifecycle Summary (31 days): family: mentions=285 active_days=26 MEMORY.md=✅(9x) trend=stable identity: mentions=342 active_days=27 MEMORY.md=✅(8x) trend=stable constraints: mentions=198 active_days=22 MEMORY.md=✅(7x) trend=rising memory: mentions=1056 active_days=30 MEMORY.md=✅(9x) trend=declining monetization: mentions=597 active_days=20 MEMORY.md=✅(3x) trend=declining content: mentions=812 active_days=28 MEMORY.md=✅(5x) trend=declining infrastructure:mentions=659 active_days=27 MEMORY.md=✅(6x) trend=declining automation: mentions=523 active_days=24 MEMORY.md=✅(4x) trend=declining research: mentions=298 active_days=15 MEMORY.md=✅(3x) trend=rising architecture: mentions=445 active_days=22 MEMORY.md=✅(5x) trend=stable