Problem
An LLM agent is stateless: every session starts from scratch, and users re-enter the same project details and preferences. LinkedIn saw this with Hiring Assistant, its AI agent for recruiters. The first memory was a key-value store (on Couchbase and Espresso) with read/write APIs only. Each application team had to extract preferences, design the index and call the API itself, so every integration was custom and adoption stayed low.
Example from the article: a recruiter archives a candidate for lacking machine learning skills. The next time that recruiter hires an ML engineer, the agent stresses that skill in outreach.
Solution in brief
The Cognitive Memory Agent (CMA) is a shared memory service with three parts: memory layers (conversational, episodic, semantic, procedural), asynchronous ingestion where an LLM turns raw activity into memory records, and a retrieval orchestrator, an LLM that plans which layers to query and in what order.
Questions this case answers
- How to give a stateful LLM agent durable per-user memory that is extracted automatically and retrieved by reasoning over several stores
- How to evaluate a memory-augmented agent separately for stored-memory quality, retrieval quality and task success
- How to make an LLM agent remember each user's preferences across sessions without asking them to state what to remember?
- How to retrieve the right memory when the answer may sit in several stores with different latency and abstraction?
Tools and methods
- Memory layers — session summary, timestamped events, generalized preferences, workflow patterns
- Memory prompts — developer-written instructions on which preferences to mine
- Hierarchical semantic memory — per-user tree of Q&A pairs, summaries and facets
- Retrieval orchestrator — LLM that plans memory tool calls and writes the answer
- Three-tier evaluation — memory quality, LLM-as-judge response quality, task success
Google or a one-time email code. No password.