Artem Getman
Current AI agents can improve within one chat, then forget everything in the next.
Core trick
Converts tool errors into structured lessons, retrieves them by task context, and promotes or suppresses each lesson by measured outcome.
Description
Current AI agents can improve within one chat, then forget everything in the next. Cortex solves
that by giving agents persistent memory that learns from failures automatically. It captures
runtime mistakes (tool errors, wrong syntax, bad action choices), turns them into structured
lessons, and retrieves the most relevant lessons by domain/task context during future runs.
Lessons are outcome-scored: helpful lessons are reused/promoted, harmful ones are suppressed. In
our mixed benchmark (gridtool -> fluxtool -> shell/excel -> sqlite -> grid retention), Wave 1
starts cold and fails multiple tasks; Waves 2–3 recover with higher pass rates, fewer errors,
and fewer steps. This moves agents from stateless chat behavior toward cumulative learning from
real execution history.
README excerpt
# Cortex: Persistent Memory For Agents Claude Code Hackathon Project Feb 2026 (Built with Opus 4.6) Current AI agents usually improve only inside one chat. New chat, same mistakes again. That breaks real productivity. Project north-star guidance is maintained in `CLAUDE.md` (and `AGENTS.md`, which symlinks to it). ## Key Findings - **Architecture trade-off:** In a 10-run-per-arm comparison, the simplified architecture beat the full design `90%` to `80%` on the easier `gridtool` task, while the full design beat simplified `100%` to `60%` on the harder `fluxtool` holdout. There was no universal winner. See [Architecture A/B Findings](docs/archive/memory-v2-history/AB-FINDINGS.md). - **Positive learning result:** On one hard shell-transfer test with 10 runs per arm, learning ON passed `5/10` versus `2/10` for OFF and reduced mean errors from `5.5` to `3.7`. See [Shell Hotfix Hard ON/OFF](tracks/cli_sqlite/reports/2026-03-09_shell_hotfix_hard_onoff_step6_10run.md). - **Negative learning result:** Under a tighter three-step protocol with a different model, learning ON passed `5/10` versus `7/10` for OFF, with low lesson activation. Memory did not help in that setup. See [Hard Transfer Benchmark ON/OFF](tracks/cli_sqlite/reports/hard_shell_git_transfer_hotfix_step3_onoff_10run.md). - **Failure diagnosis:** Controlled tests separated insufficient step budget, inactive memory retrieval, and model capability as distinct bottlenecks. See [Three-Part Diagnosis](tracks/cli_sqlite