Most “memory” demos are really just chat history. Useful, but it’s not the same as remembering things across sessions in a way that’s safe and intentional. True agent memory is a product feature that requires careful design.
Chat history is not memory
Appending every message to a context window is the simplest form of memory, and the least useful at scale. It burns tokens, dilutes relevance, and creates privacy problems when users expect that earlier conversations stay private.
Real memory is selective. It stores what matters, forgets what doesn’t, and retrieves the right context at the right time. This is the difference between an agent that repeats itself and an agent that genuinely learns.
Start with scope
Good memory starts with scope. Cortex handles this for Teleon-powered agents with three distinct layers.
The company layer holds shared knowledge policies, FAQs, product documentation. Every agent in your organization can access it. The team layer holds context specific to a group playbooks, shared notes, team-specific workflows. The personal layer holds user-specific information preferences, past interactions, ongoing projects.
Scope enforcement means an agent serving User A never accidentally surfaces User B’s data. In Cortex, this is enforced automatically through scope fields like `user_id` and `team_id` on every memory operation.
Decide what to remember and what to forget
The most common failure mode is over-remembering. If everything is stored forever, you’ll leak context, confuse the model, and create privacy headaches.
Before adding memory to an agent, answer three questions: What should this agent remember between sessions? How long should that memory persist? And under what conditions should it be deleted?
For a customer support agent, you might want to remember the customer’s plan, recent tickets, and stated preferences. You probably don’t want to remember every individual message. For a coding assistant, you might want to remember project structure and conventions, but forget debugging sessions once the issue is resolved.
Retrieval matters as much as storage
Storing memories is only half the problem. Retrieving the right ones at the right time is what makes memory useful. Cortex provides two retrieval strategies: recency-based (recent history) and semantic search (relevance-based).
Before your agent runs, Cortex automatically retrieves recent interactions and semantically relevant memories, then injects them into the agent’s context via `cortex.context`. This means your agent always has the right background without you manually querying a database.
Memory as a product differentiator
Agents with well-designed memory feel qualitatively different to users. They don’t ask the same questions twice. They pick up where the last session left off. They get better over time. This isn’t magic, it’s careful engineering applied to what gets stored, how long it persists, and when it surfaces. Explore how Cortex implements this, or check out the features page for the full picture.