> For the complete documentation index, see [llms.txt](https://docs.questera.ai/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://docs.questera.ai/questera-ai-platform/ai-agents/agent-memory.md).

# Agent Memory

1\. Three-Tiered Memory Architecture

a. Short-Term Memory (Session or Campaign-level)

* Stores recent user actions (e.g., "clicked last email", "ignored upsell offer").
* Lives for the current session or short lifecycle context (e.g., 7–14 days).
* Enables reactive personalization (e.g., don’t re-offer something just clicked).

Example:\
Churn agent recalls that the user clicked "cancel reasons" → triggers a save offer, not generic re-engagement.

b. Long-Term Memory (User-level + Segment-level)

* Captures patterns across lifecycle stages: onboarding, activation, conversion, churn.
* Includes vectorized summaries of:
* Past interactions
* Channel preferences
* Campaign responses
* Agent actions + outcomes

Stored in a Vector DB (e.g., Weaviate or Pinecone) using user embeddings or segment-level embeddings for fast semantic retrieval.

Example:\
For a mid-funnel B2B user, the agent recalls that whitepapers → demo → closed won is a high-conversion path and prioritizes this path again.

c. Global Memory (Agent + Team-level)

* Shared knowledge across all agents and customers (pseudonymized).
* Stores best-performing prompts, subject lines, CTAs, channel combos.
* Updated with reward signals (e.g., lift in conversions, reduced drop-offs).

This enables self-improvement and cross-agent knowledge sharing (e.g., churn prevention agent learns from upsell agent’s success).

2\. How Memory Is Queried

Agents use memory via:

* Embedding-based retrieval (vector similarity)
* Metadata filtering (e.g., campaign type = “churn save”)
* Time decay logic to prioritize fresh interactions

Often wrapped in a Retriever module inside LangChain / LangGraph agent flows.

3\. What Gets Stored

* User & segment interaction logs
* Campaign variants tried
* Reasoning steps (in CoT, ReAct form)
* Final outcomes (clicks, conversions, revenue delta)
* Feedback from marketers (human-in-the-loop scoring)

4\. Memory Refresh & Pruning

* Low-signal events decay over time
* Abandoned flows are eventually purged
* Only high-relevance, high-reward outcomes are retained long-term

<br>