Total Recall: Agent Memory and Harness Engineering — Ignacio Martinez, Oracle
Treat memory, context control, concurrency, and model routing as harness concerns. The workshop shows how these layers can improve agent behavior without changing model weights.
The workshop separates the model’s fixed reasoning core from the harness around it. It covers **agent memory**, combining files with databases, using worktrees when many agents edit concurrently, and routing tasks between models.
Builders should keep working context deliberately small, persist reusable knowledge outside the conversation, and turn effective sessions into repeatable workflows. For parallel coding agents, isolate file changes rather than expecting shared files to provide transactional consistency.
The workshop separates the model’s fixed reasoning core from the harness around it. It covers **agent memory**, combining files with databases, using worktrees when many agents edit concurrently, and routing tasks between models. Builders should keep working context deliberately small, persist reusable knowledge outside the conversation, and turn effective sessions into repeatable workflows. For parallel coding agents, isolate file changes rather than expecting shared files to provide transactional consistency. The session presents these as practical design patterns, not measured performance results. Claims about context degradation and a future **mixture of small experts** are architectural guidance; the material provides no benchmark against simpler harnesses.
This consolidates external memory, compact working context, model routing, and worktree isolation into one practical harness design. It reinforces durable-state and compound-engineering patterns, but offers no comparative evidence that the full bundle beats simpler scaffolds; the empirical candidates further narrow its prescriptions by showing that context controls and tools should vary with model and budget.