Enterprise Agentic AI Memory Lab (Open Source)

Author: Satish Gopinathan Web: https://www.eagleeyethinker.com

This project demonstrates five memory types in one agentic HR Policy Assistant.

Memory Types

1. Short-Term Memory -> Redis (session)
2. Retrieval Memory -> ChromaDB + Ollama embeddings
3. Semantic Memory -> NetworkX graph
4. Procedural Memory -> LangGraph workflow
5. Durable Memory -> JSON audit log

Requirements

• Python virtual environment with dependencies:
  • pip install -r requirements.txt
• Ollama running locally with model:
  • ollama pull llama3
• Redis-compatible server running on localhost:6379:
  • Redis or Memurai

Install Redis

Windows (Memurai, Redis-compatible)

choco install memurai-developer -y
Start-Service Memurai

Optional connectivity check:

& "C:\Program Files\Memurai\memurai-cli.exe" ping

macOS (Redis via Homebrew)

brew install redis
brew services start redis

Optional connectivity check:

redis-cli ping

Run

python app/main.py

What a New Developer Should Expect

1) Short-Term Memory (Redis session)

• Each prompt you type is appended to a Redis list under session key default.
• The session key TTL is set to 1800 seconds (30 minutes) on each new input.
• Expected behavior:
  • Recent user turns are available during the active session.
  • If Redis is down, app calls to short-term memory will fail.

Quick check (if using Memurai):

& "C:\Program Files\Memurai\memurai-cli.exe" LRANGE default 0 -1
& "C:\Program Files\Memurai\memurai-cli.exe" TTL default

2) Retrieval Memory (ChromaDB + Ollama)

• On startup, data/policies.txt is loaded and split into chunks.
• Chunks are embedded with llama3 embeddings and indexed in ChromaDB.
• For each query, top k=2 similar chunks are retrieved and used as LLM context.
• Expected behavior:
  • Answers should be grounded in policy text when relevant.
  • If Ollama is not reachable, app startup/query will fail.

3) Semantic Memory (NetworkX)

• A small in-memory concept graph is created with fixed relationships:
  • Employee <-> Manager
  • Manager <-> Director
  • Policy <-> Compliance
• For every query, neighbors of Policy are fetched.
• Expected behavior:
  • Semantic relation list currently resolves to ["Compliance"] and is logged.

4) Procedural Memory (LangGraph)

• The workflow is a single-node LangGraph pipeline:
  • input -> agent node -> END
• Agent node order:
  1. Store input in short-term memory
  2. Retrieve relevant policy chunks
  3. Read semantic relationships
  4. Build prompt and call LLM
  5. Write durable audit record
• Expected behavior:
  • Every user prompt consistently follows the same execution path.

5) Durable Memory (JSON audit log)

• Every turn is appended to audit_log.json.
• Each record contains:
  • input
  • retrieval (retrieved chunks)
  • semantic (graph relationships)
  • output (model response)
• Expected behavior:
  • audit_log.json grows over time and preserves past runs for inspection.

Use Case

HR Policy Assistant