Glossary
Typed memory (AI agent)
Typed memory in an AI agent is a structured memory system where every stored piece of information is attached to a precise semantic type (e.g., brand_voice, audience, recent_wins, user_prefs, feedback) rather than being dumped as unstructured free text.
Also known as
- typed memory
- agent memory
- structured memory
Typed memory solves the 3 core problems of free-text memory: (1) **precise retrieval** — you can query only the voice block when generating a post, without noise from everything else; (2) **controlled editing** — the user can edit their audience block independently; (3) **conflict management** — when the post-turn consolidator detects that a new fact contradicts an older one, it can perform typed supersession (the new replaces the old, the old moves to history with a timestamp) rather than a naive append that ends in schizophrenic memory.
2026 architectures add: per-block confidence (how certain the agent is about a given fact), provenance (which conversational turn produced the information), and correction-triggered invalidation (if the user says "that's wrong," the agent knows exactly which block to invalidate).
In the getchatsocial.com product
getchatsocial.com uses 5 block types: brand_voice, audience, recent_wins, user_prefs, feedback. A post-turn consolidator (cheap Gemini-flash) extracts learnings from each conversation with typed supersession and confidence scoring. Visible and editable on the /memory page.
FAQ
What's the difference from ChatGPT Memory?
ChatGPT Memory is an unstructured text dump. Typed memory organizes information by category and allows targeted editing and deletion. For marketing use cases, that level of control is critical.
How does it learn?
After each turn, a consolidator (typically a cheap LLM) extracts new learnings with their type and confidence, then compares them against existing blocks. If there's a conflict, typed supersession kicks in. If compatible, the blocks are merged.