Fetching from the wire…
01
02
03
04
05
06
07
08
Public story · 2026-08-25 · high
Four memory layers, conversational, episodic, procedural and semantic, run through a three-tier check for context pollution before reaching recruiters.
Why now: Bodigutla laid out the architecture in a Stack Overflow interview published August 25.
LinkedIn holds its hiring agent's memory system to 10 to 20 percent of total response time, per Principal AI Researcher Praveen Bodigutla. He laid out the architecture in a Stack Overflow interview. Recruiters move through candidates fast, so every extra second the memory layer spends competes directly with the matching work itself.
Bodigutla says the system stays inside that budget through three moves. It runs memory retrieval in parallel with planning, calls the LLM only when needed, and caches prefixes instead of recomputing them on every turn.
The four layers split by what they track. Conversational memory covers the current session. Episodic memory holds a timeline of past interactions with provenance back to the source conversation. That lets the agent point to where a preference came from. Procedural memory captures how a specific recruiter trades off criteria like years of experience against skills match. Semantic memory rolls those patterns up into a standing preference profile across sessions and surfaces.
LinkedIn also moved off GraphRAG to a tree-structured memory that updates incrementally, without rebuilding a graph on every write. A three-tier evaluation checks that the system preserves entities and cites where each fact came from. It's meant to catch context pollution, one bad memory poisoning later answers, before it reaches a recruiter.
Swapping GraphRAG for a tree structure to hit a latency target says something about the field beyond LinkedIn. Precision-first retrieval designs are losing to speed once agent memory has to run inside a live product's latency budget. Watch whether other companies shipping memory-heavy agents make the same trade.
Each link below shares sources, entities, or timing with this story.
Simon Willison released LLM / Shared entity: LLM / Shared topic / Earlier coverage / Tension
Linked by a graph relationship (Simon Willison released LLM); both cover LLM; overlapping topics (agent, call).
Microsoft released GraphRAG / Shared entity: LLM / Shared topic / Earlier coverage
Linked by a graph relationship (Microsoft released GraphRAG); both cover LLM; overlapping topics (agent, context).
Simon Willison released LLM / Shared entity: LLM / Shared topic / Earlier coverage
Linked by a graph relationship (Simon Willison released LLM); both cover LLM; overlapping topics (agent, call).
Stack Overflow partners with OpenAI / Shared entity: LLM / Shared topic / Earlier coverage
Linked by a graph relationship (Stack Overflow partners with OpenAI); both cover LLM; overlapping topics (caching, call).
Stack Overflow partners with OpenAI / Shared entity: LLM / Earlier coverage / Tension
Linked by a graph relationship (Stack Overflow partners with OpenAI); both cover LLM; earlier LLM coverage from 2026-07-27.
Microsoft released GraphRAG / Shared entity: LLM / Earlier coverage / Tension
Linked by a graph relationship (Microsoft released GraphRAG); both cover LLM; earlier LLM coverage from 2026-07-26.
Simon Willison released LLM / Shared entity: LLM / Earlier coverage / Tension
Linked by a graph relationship (Simon Willison released LLM); both cover LLM; earlier LLM coverage from 2026-06-19.
Linked by a graph relationship (Simon Willison released LLM); both cover LLM; earlier LLM coverage from 2026-06-18.