Awesome Search KG

AI agent memory: Creating smart agents with Elasticsearch managed memory

3 min read

AI agent memory: Creating smart agents with Elasticsearch managed memory

Source: https://www.elastic.co/search-labs/blog/ai-agent-memory-management-elasticsearch Published: 18 March 2026 · Authors: Gustavo Llermaly, Jeffrey Rengifo (Elastic Search Labs)

Summary

A working implementation of Agentic Memory on Elasticsearch, organised around a three-way memory taxonomy and — the part that distinguishes it — using document-level security as the memory isolation mechanism rather than application-level filtering.

The Three Memory Types

TypeWhat it holdsWhere it lives
ProceduralHow the agent behaves — when to store a memory, when to retrieve one, how to summarize, how to use toolsApplication code and prompts; not stored in Elasticsearch
EpisodicSpecific experiences tied to an entity and a context — “Peter’s birthday is tomorrow and he wants steak”Documents in Elasticsearch, with user, role, timestamp and context metadata
SemanticAbstracted, generalized world knowledge independent of any single interaction — a company handbookRetrieved through the same system, with a different retrieval strategy

The article singles out episodic memory as “the most dynamic and personal” and the one “most prone to context pollution if handled incorrectly” — which is what motivates the isolation work below.

Memory Isolation via Document-Level Security

The demo is framed on the TV show Severance: one agent, “Mark”, whose work-self (“innie”) and personal-self (“outie”) memories must never mix.

Rather than filtering in the agent, each user’s credentials carry a role whose descriptor pins a filter onto the memory_type field:

"query": {"bool": {"filter": [{"term": {"memory_type": "innie"}}]}}

The outie role is identical but for "memory_type": "outie". Elasticsearch then enforces the separation itself — the agent issues the same query regardless, and the engine returns only what those credentials permit. Cross-context access requires a separate user or an additional role, which the article leaves as an exercise.

Index and Retrieval

Memory documents carry:

  • user_id — keyword
  • memory_type — keyword (the security discriminator)
  • created_at — date
  • memory_text — text, with a multi-field semantic subfield

The multi-field setup indexes the same content twice, for full-text search and for semantic search via ELSER. Retrieval is hybrid, combining the two with RRF at rank_window_size: 50 and rank_constant: 20.

Agent Tools

Three functions are exposed to the model, which chooses among them via OpenAI’s Response API function calling:

  • GetKnowledge — conventional RAG-style context search
  • GetMemories — hybrid semantic + keyword retrieval over memories, using the RRF retriever above
  • SetMemory — writes a new memory after the LLM converts the exchange into a structured record

Results are injected into the conversation context before the final response is generated.

Caveats

Memory compression is named as an open problem and left unimplemented. Selective filtering is credited with reducing context pollution and yielding lower latency and token usage; no scale, latency or dimensionality figures are given.

Graph View

Backlinks