Skip to content

MemoryMesh Agent — User Guide

How to actually use MemoryMesh Agent once it's running. If you haven't set it up yet, start with the Setup Guide.

MemoryMesh Agent is a multi-agent market-surveillance assistant. You ask it questions in plain English; a team of specialist agents investigates, reasons over the results, and answers — and everything they learn is written into CockroachDB so future questions benefit from it.

The app shell

Open the app (http://localhost:5173 in local dev) and you'll see a left nav rail with two views: Chat and Dashboard. The header always shows two status badges:

  • Local (amber) or AgentCore (indigo) — which backend is answering your questions right now
  • CockroachDB connected (green) or CockroachDB offline (red) — live connectivity to the memory layer

If the CockroachDB badge is ever red, nothing else in the app will work correctly — see the Setup Guide's troubleshooting section.

Chat view

This is the main view: a conversation with the multi-agent system, plus a memory sidebar on the right.

Asking a question

Type a question into the composer at the bottom and send it. A few things happen, visibly:

  1. The system checks CockroachDB's long-term case memory for anything similar it has investigated before (you won't see a separate step for this — it happens before the answer starts streaming, and it changes what the agents say if a related case exists).
  2. The right specialist agent (or agents) run and stream their answer back token by token.
  3. Under the assistant's message, a compact agent pipeline graph appears — a small node graph showing exactly which agents ran and how many tool calls each one made, built live from the same event stream as the answer itself. This is your window into "who actually answered this" without reading raw logs.
  4. If an agent's investigation returns citations (see Triage citations below), they render as clickable chips under the pipeline.

Which agent handles what

You don't pick an agent — the orchestrator routes your question based on its content. Rough guide to what triggers what:

Ask about... Routes to
Trading activity for one security, one day security_monitor
One security across multiple days broker_monitor
Broker activity on a given day risk_monitor
External market news/context intel_analyst
The system's own memory — case counts, past investigations, cluster health memory_ops
Whether a finding is a compliance issue compliance_officer
How urgent an investigation is case_triage
A summary of this session's history audit_reviewer

A single question can route to more than one agent — the pipeline graph shows you exactly which ones ran.

Example queries to try

# A first investigation — this writes a new case into memory
What was the trading activity for AAPL on March 15, 2024 and which brokers were most active?

# Ask something similar later, in a new session — watch the answer's
# context include a recalled prior case
Which brokers were most active trading AAPL in mid-March 2024?

# Memory introspection — routes to memory_ops, which queries CockroachDB
# live via the Cloud Managed MCP Server
How many past investigations do we have stored, and have we looked at broker risk on MSFT before?

# Multi-agent
Analyze MSFT's price movement and broker risk scores between March 10-15, 2024, and check for any related market news.

# Compliance + triage — case_triage explicitly queries case memory to set priority
Is ALPHA_CAPITAL's AAPL trading on March 15, 2024 a compliance issue, and how urgent is it?

# Audit — reads message_store directly, not the model's own recollection
Give me an audit summary of everything asked and answered in this session.

Triage citations

When case_triage runs, it calls recall_similar_investigations against the vector index itself (a second, agent-driven path into the same memory the automatic recall step uses). If it finds matches, they show up as small citation chips — case ID and similarity score — directly under the pipeline graph, so you can see exactly which past cases informed the urgency assessment instead of taking the model's word for it.

History — replaying a session's checkpoints

Click History in the header (next to your session ID) to open the time-travel view: a scrubber over every checkpoint LangGraph has saved for the current session, in order. Each step shows the actual workflow state at that point — which node had just run, what it had produced. Because AsyncCockroachDBSaver already persists a checkpoint after every node, this view costs nothing extra to build — it's a live read of state that was already being saved.

Memory sidebar

The right-hand panel (desktop only) shows three live counters — Cases, Turns, Sessions — polled every few seconds directly from CockroachDB, plus a short list of recent cases. Watch the case count tick up as you chat; each finished investigation writes itself back into case_memory, so the count is real accumulated history, not a per-session tally.

New session

Click New session to start a fresh conversation. Past sessions aren't deleted — they're still in message_store and queryable by audit_reviewer, and any cases they produced are still in long-term memory for future recall.

Dashboard view

Switch to Dashboard in the nav rail for an operational view of the same CockroachDB tables the agents use — everything here is a live read, not a mock.

  • Stat tiles — cases, turns, sessions, and system health at a glance.
  • Agent activity chart — a bar chart of tool calls per agent, colored consistently (an agent always gets the same color, regardless of how the bars are currently sorted — earlier versions of this chart had a bug where colors were assigned by sorted rank and would visibly shift between polls; that's fixed).
  • Cases-per-day chart — a time series of investigation volume over the last 14 days.
  • Vector Memory Map — every case's embedding, projected to 2D and plotted as a point cloud. Type a query into its search box and it runs the same similarity search the agents use internally, plots your query alongside its real nearest neighbors on the same projection, and lists the ranked matches below. This is the distributed vector index made visible and queryable, not just a number.
  • Recent cases table — the latest investigations, most recent first.
  • System health — backend mode and CockroachDB connectivity, same badges as the header.

Command-line alternative

Prefer a terminal over the browser? python scripts/chat_cli.py runs the exact same LangGraph workflow with no HTTP layer at all — useful for scripting or quick checks.

Tips

  • Ask a near-duplicate question in a different session to see recall working — the giveaway is the answer referencing a prior finding you never mentioned in the current conversation.
  • If you want the Dashboard to have something to show immediately on a fresh cluster, run make seed-memory before you start chatting (see the Setup Guide).
  • The MCP-powered memory_ops agent only works if COCKROACHDB_MCP_API_KEY is set; without it, questions about the memory cluster itself will get a graceful "can't introspect the live cluster" answer instead of an error.