Chat history is not agent memory: production lessons on adaptive agents
_jaydeepkarale · x · 2026-09-03
- Core argument: most agent implementations store every message and dump the full conversation into the next prompt. That's wrong — conversations balloon, tokens drive up latency, and you end up doing gymnastics to fit the context window.
- Beyond size, key failure modes: noisy context (old tool results and stale facts compete with useful info), lost continuity across sessions, incorrect retrieval (similarity search returning the wrong user/thread/workflow), and conflicting memories.
- Accompanied by the DeepLearning.AI short course "Building Adaptive AI Agents" (with Oracle): skill induction from agent traces with human approval, a code knowledge graph for context retrieval, and three adaptation layers (behavior, knowledge, model).
More from coding & agent
- Cohere Labs releases ATE dataset of ~700K MCP tools to reveal what agents do — Cohere_Labs · 2026-09-03
- The Mythical Agent Month: past a threshold, every new agent adds more work than it removes — viksit · 2026-09-03
- ngrok launches AI Gateway: route local LLMs and OpenAI through one URL — anthara_ai · 2026-09-03
- Google ships agent-specific access settings — apps need to get agent-friendly — 0xsachi · 2026-09-03
- Frontier Models Give Suspiciously Bad Advice on Local AI Setups, Developer Reports — ikilaie · 2026-09-03
- Two engineers rebuilt a FedEx-scale delivery system core in 3.5 months with AI — alex_verem · 2026-09-03