Selmem experiments show divergent memories don't yield divergent behavior without a self-model
Tough_Conflict_8852 · reddit · 2026-11-05
Developer jbsalles had GPT write a narrative from the first 100 commits of Selmem, a selective, reconstructive memory system for LLMs, tracing how experiments reshaped the initial hypothesis: two identical LLMs with different experiences and memories should diverge — but divergent memories don't mean divergent behavior, and influencing behavior doesn't mean learning something useful.
Six stages: (1) Sept: building living memory with selection, forgetting, reconstruction, sleep, consolidation; (2) late Sept: memories diverge but weakly influence decisions; (3) early Oct: making behavior measurable via causal ablations and branching experiments; (4) mid Oct: an evolving self-model becomes necessary to guide memory selection; (5) late Oct: measurable influence via contextual beliefs and self-model-guided retrieval; (6) Nov: moving toward alignment — preserving rules and their exceptions.
Key architectural decision: separating a stable semantic core, a mutable gist, and a sealed audit-only archive so recall can't rewrite history. The author notes an irreducible tension between grounding (protecting memory fidelity) and transformation (enabling individuality). Early-stage personal project with 11 GitHub stars.
More from coding & agent
- Musk shows off Grok calling a dentist to check insurance coverage — elonmusk · 2026-11-05
- Treat AI bots as employees, not assistants: an onboarding playbook — elonmusk · 2026-11-05
- Codex Voice cuts tool latency by streaming context to the harness before delegation — SIGKITTEN · 2026-11-05
- I split AI short-drama production into 5 agents—here's what broke — LateQuantity4194 · 2026-11-05
- Keynote built entirely with OpenClaw agent via voice notes on a train ride — vincent_koc · 2026-11-05
- DeepSeek's dsh-superagent: a coding agent whose rulebook is 'stop narrating your work' — shoumikgoswami · 2026-11-05