itamar-friedman, CEO of Qodo (AI code review, quality, and testing, “the counterpart to vibe coding”), discusses memory for coding agents. Short-term memory keeps a long session faithful to early requests; long-term memory is project, personal, and organizational knowledge. Rules in markdown files are fine, since natural language keeps memory controllable, but you need a framework: how memories are created, when they’re retrieved, and whether they’re valuable. Track rule usage and acceptance (e.g. used 100 times, accepted 80) and prune. Ingestion design: between “encode everything into one format” and “only index references”, ingest and encode per media type with references back to sources, give agents per-type memory tools, and add a router. Governance (whose feedback counts: junior, senior, an outlier senior?) is what’s missing in tools. Mem0 offers add, search, update, delete, plus multimodal, graph, expiry, and feedback, but not the intelligence. Merging parallel sessions needs their intent, not just their code. Future: nightly background consolidation (like sleep) plus backfilling history to avoid cold start; possibly learned memory architectures, despite the bitter lesson.
Show: langtalks
Concepts: agent-memory · memory-consolidation · human-vs-ai-memory · context-engineering