Perseus Memory Agent — Elastic Partner Track

Your AI agent shouldn't have amnesia. Perseus gives Gemini agents persistent memory across sessions remembering your stack, decisions, and lessons. Elastic (cloud) or Engram-rs (self-hosted) same API.

Comment

What It Does

Perseus Memory Agent gives AI agents persistent, evolving memory so they remember project context across sessions. Instead of re-explaining your tech stack, conventions, and architectural decisions every time you start a new session, the agent recalls everything it learned about your project — even from weeks ago.

Key capabilities:

  • Remembers project context — stack, conventions, architecture, preferences
  • Logs decisions with rationale — why pgvector over Pinecone? The agent remembers
  • Compounds knowledge — spots patterns across sessions, surfaces insights
  • Swappable backends — Elastic (managed cloud) or Engram-rs (self-hosted MIT), one config line
  • Hybrid search — ELSER semantic + BM25 keyword for accurate recall via Elastic

The demo shows 3 sessions:

  1. Session 1 — agent learns project context from scratch
  2. Session 2 — agent recalls prior knowledge, logs decisions with rationale
  3. Session 3 — agent compounds everything, backend swap Elastic → Engram-rs

How I Built It

Built with Gemini 3 Pro in Google Cloud Agent Builder, connected to Elastic Agent Builder via MCP for the memory layer:

  1. Elastic MCP tools — Three tools in Elastic Agent Builder: search_memory (hybrid ELSER + BM25), store_memory, delete_memory. Exposed as MCP endpoints the Gemini agent calls directly.

  2. Memory abstraction layer — Abstract MemoryBackend interface in Python. Two implementations: ElasticMemoryBackend (Elasticsearch via MCP) and EngramMemoryBackend (Engram-rs CLI / SQLite). Same API surface, swap backends by changing one environment variable.

  3. Agent tools — Three MCP-callable tools: ProjectContextTool (manage project stack/conventions), DecisionLogTool (log and recall architectural decisions with rationale), KnowledgeGraphTool (cross-reference memories, find patterns, compound knowledge).

  4. Session lifecycle — start_session() loads all relevant context from Elastic. process_message() enriches every prompt with hybrid search results. end_session() reflects on new knowledge and compounds insights.

  5. Backend-agnostic — The abstract interface means the same agent code works with Elastic (managed) or Engram-rs (self-hosted). Demo shows the swap live.

Why Elastic

Elastic Agent Builder was the natural choice because it's the only partner with a built-in context layer for memory and insights:

  • Hybrid search (ELSER semantic + BM25 keyword) gives accurate recall that pure vector search or pure keyword search can't match alone
  • ES|QL enables complex memory queries without custom code
  • MCP server makes tools immediately available to Gemini — no custom integration layer needed
  • Elasticsearch scales from hackathon demo to production without changing the API

What's Next

  • Multi-project awareness — Agent recognizes cross-project patterns (e.g., "you use this same auth pattern in 3 repos")
  • Memory confidence decay — Facts automatically lose confidence if unverified, prompting re-verification
  • MCP-native Engram-rs — Direct MCP server in Engram-rs so it plugs into Agent Builder without the CLI wrapper

Built With

  • Google Cloud Agent Builder (Gemini 3 Pro)
  • Elastic Cloud (Elasticsearch, ELSER, BM25)
  • Elastic Agent Builder (MCP tools)
  • Engram-rs (Self-hosted memory, MIT, Rust + SQLite + FTS5)
  • Python 3.12 (Pydantic, async/await)
  • MCP (Model Context Protocol)

Built With

Share this project:

Updates