Long-term memory
The agent's memory that survives across sessions, sometimes across months or years — usually a vector store plus a key-value store, with episodic, semantic, and procedural layers underneath.
Long-term memory is the umbrella that contains episodic (events), semantic (facts), and procedural (skills) memory. It is what makes an agent useful on session 200 not just session 1.
In 2026 the typical architecture: vector store for semantic retrieval of past summaries, key-value store for high-confidence facts, and a small graph for relationships. Vendors like Letta, Mem0, Zep, and Pinecone Memory wrap this stack.
The hard parts of long-term memory are eviction, conflict resolution, and privacy. Models will happily remember everything; the question is what to forget, what to update, and what to never have written down in the first place.
Frequently asked
How big should an agent's long-term memory get?+
Most production agents converge on 5–50 MB of summarized state per active user. Raw chat history dwarfs this but is mostly redundant — summaries dominate quality.
Should I let users see what the agent remembers?+
Yes. The best products in 2026 ship a "memory panel" where users can read, edit, and delete the facts the agent has on file. Trust is downstream of transparency.
