Recruiting · Production

LinkedIn builds Cognitive Memory Agent to enable personalized, stateful AI agents at scale

The problem

LinkedIn's AI agents needed persistent memory to learn from interactions and deliver personalized experiences, but traditional memory systems required users to explicitly specify what to remember, and an earlier in-house experiential memory store provided infrastructure with no intelligence, forcing every application team to build custom extraction and retrieval logic.

First attempt

LinkedIn's initial experiential memory was a hierarchical key-value store built atop Couchbase and Espresso with no intelligence; each application team had to manually build preference extraction, indexing, and retrieval logic, making every integration custom-built and preventing ubiquitous adoption.

Workflow diagram · grounded in source
1
User activity data ingested
integration
“Raw user activity and conversation data must be indexed into various memory formats before being stored, enhancing retrieval efficiency. For this, we offer two pathways: streaming and batch.”
2
LLM memory extraction
ai_action
“This process involves using LLMs to summarize learnt patterns for long-term memory, extract episodic activities specified in memory prompts and periodically compress conversational memory”
3
Natural language query received
trigger
“The CMA orchestrator receives the natural language query, relevant memory keys, and the available memory tools.”
4
Retrieval plan generated
ai_action
“It then generates a retrieval plan. For simple queries, this plan may involve a single semantic lookup. For more complex questions, the orchestrator may sequentially invoke multiple memory providers, refine intermediate queries, and synt…”
5
Multi-layer memory retrieval
ai_action
“the orchestrator may first consult aggregated semantic memory to understand inferred recruiter preferences and long term hiring patterns. It may then query episodic memory for recent positive feedback on non-identifiable features of simi…”
6
Personalized suggestion delivered
output
“CMA is leveraged by the Hiring Assistant to fetch the recruiter's past projects and preferences to suggest a role or title that they can start hiring for. Typically recruiters hire for similar roles in a company, in such cases CMA provid…”
Reported outcome

LinkedIn built the Cognitive Memory Agent (CMA), a horizontal memory platform powering stateful, context-aware AI agents at scale, now applied in the globally available Hiring Assistant to personalize recruiter interactions and reduce friction by automatically surfacing past preferences.

Reported metrics
Recruiter productivityproductivity boost
User frictionreduces user friction
Reported stack
Cognitive Memory AgentHiring AssistantCouchbaseEspressoopen-source LLMvector store
Source
https://www.linkedin.com/blog/engineering/ai/the-linkedin-generative-ai-application-tech-stack-personalization-with-cognitive-memory-agent
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

LinkedIn built the Cognitive Memory Agent (CMA), a horizontal memory platform powering stateful, context-aware AI agents at scale, now applied in the globally available Hiring Assistant to personalize recruiter intera…

What tools did this team use?

Cognitive Memory Agent, Hiring Assistant, Couchbase, Espresso, open-source LLM, vector store.

What results were reported?

Recruiter productivity: productivity boost; User friction: reduces user friction (source-reported, not independently verified).

What failed first in this deployment?

LinkedIn's initial experiential memory was a hierarchical key-value store built atop Couchbase and Espresso with no intelligence; each application team had to manually build preference extraction, indexing, and retrie…

How is this recruiting AI workflow structured?

User activity data ingested → LLM memory extraction → Natural language query received → Retrieval plan generated → Multi-layer memory retrieval → Personalized suggestion delivered.

Related recruiting cases