2BAgent — Interactive Demo & Architecture

Feature walkthroughs

See every layer, step by step

Each tab shows a realistic trace of what happens when you use that feature — from the command you type to what the system does internally.

You type a thought in Discord. Here is exactly what happens in the next 200ms.

💬

Discord

/capture command

→

🌐

FastAPI

POST / · port 8900

→

🧬

Extractor

type · people · topics

→

🤖

Ollama

1024-dim embedding

→

🗄️

PostgreSQL

thoughts + pgvector

Discord → API

ram today at 9:14 AM

/capture DECISION: use asyncpg over SQLAlchemy — raw SQL gives us full control over query planning and index hints

2BAgent APPtoday at 9:14 AM

✅ Captured

Type: decision
Topics: #database #architecture
People: —
Stored · embedding 1024d · 143ms

What gets stored in PostgreSQL

thoughts table — live row

id: "a3f8c2d1-..."
raw_text: "DECISION: use asyncpg..."
source: "discord"
metadata: {
"type": "decision",
"topics": ["database","architecture"],
"people": []
}
embedding: vector(1024) [0.021, -0.094, ...]
fts_vector: tsvector (GIN indexed)
created_at: "2026-06-16T09:14:22Z"

Supported prefixes (case-insensitive):
DECISION: insight: ToDo: todo- task: meeting: PERSON: — or plain text → auto-classified as note

Hybrid search combines vector cosine similarity with PostgreSQL full-text search. Here's a real query trace — including what happens when your keywords don't match what you wrote.

Discord query

ram today at 2:31 PM

/search what database decisions did I make?

2BAgent APPtoday at 2:31 PM

🔍 4 results · hybrid mode

0.94 — DECISION: use asyncpg over SQLAlchemy...
0.91 — DECISION: pgvector cosine over Euclidean...
0.87 — INSIGHT: asyncpg bottleneck on high concurrency...
0.72 — NOTE: looked into pgbouncer connection pooling...

SQL executed (hybrid)

GET /search?q=database+decisions&mode=hybrid

-- Step 1: embed query via Ollama
-- "what database decisions did I make?"
-- → 1024-dim vector q_emb

-- Step 2: set vector probe depth
SET ivfflat.probes = 10;

-- Step 3: hybrid merge query
SELECT id, raw_text, metadata,
(0.7 * vec_score + 0.3 * fts_score)
AS score
FROM (
SELECT *, 1 - (embedding <=> $1::vector)
AS vec_score,
ts_rank(fts_vector, q) AS fts_score
...
) ORDER BY score DESC LIMIT 10;

Result details — ranked by hybrid score

0.94

decision · 3 weeks ago · discord

DECISION: use asyncpg over SQLAlchemy — raw SQL gives us full control over query planning and index hints

decision#database#architecture

vec: 0.96 · fts: 0.88 · merged: 0.94

0.91

decision · 5 weeks ago · mcp

DECISION: pgvector cosine distance over Euclidean — cosine is unit-invariant, handles varied embedding magnitudes correctly

decision#pgvector

vec: 0.93 · fts: 0.85 · merged: 0.91

0.87

insight · 2 weeks ago · discord

INSIGHT: asyncpg connection bottleneck under 50+ concurrent writes — need to move to a pool before load testing

insight#database

vec: 0.88 · fts: 0.84 · merged: 0.87

0.72

note · 4 days ago · api

looked into pgbouncer for connection pooling — might reduce latency under the reconcile loop concurrency issue

note#database

vec: 0.71 · fts: 0.74 · merged: 0.72

Why this works: You searched "database decisions" — not "asyncpg" or "pgvector". The vector similarity found these because the embeddings captured the conceptual meaning, not just the keywords. The third result uses "bottleneck" not "decision" — found by meaning, not text match.

Every morning at 7:00 AM the agent synthesizes your last 24 hours of captures, open tasks, calendar events, and recent emails into a single briefing — then posts it to your Discord agent channel.

Sample output — Discord

🌅

Good morning, Ram — here's your Monday

2BAgent · morning briefing · Mon Jun 16, 7:00 AM · via Gemini

Yesterday's captures

• 9 thoughts captured — 2 decisions, 3 insights, 3 tasks, 1 note

• Most active topics: #database (4×), #architecture (3×), #product (2×)

✓ Build habit: marked complete at 4:12 PM

Open tasks 3 open

⚠ "Review PR #47 for asyncpg pool changes" — 4 days old stale

⚠ "Follow up with Alice re: Q2 planning" — 7 days old stale

• "Migrate legacy cron jobs to agent runner" — added yesterday

Today's calendar

• 10:00 AM — Engineering standup (30 min) · Alice, Ben, Carol

• 2:00 PM — Q2 planning · Alice confirmed · agenda not set

⚠ 6 unread emails since Friday · 2 from Alice (re: Q2)

Patterns & connections

↗ "database" has appeared in 11 of your last 14 thoughts — this is becoming a major focus area. Consider a decision log.

↗ "Alice" appears in 3 open tasks and 2 calendar events today. High-leverage day for that relationship.

💡 Your decision on asyncpg 3 weeks ago is now showing up as a follow-on insight. The pattern holds.

How the briefing is generated

🧠

Think

Gather context from all sources

Agent fetches last 24h thoughts from DB, reads Gmail inbox (OAuth2), pulls today's calendar events, checks HABITS.md state, loads MEMORY.md for context. All data is assembled into the synthesis prompt.

list_recent(24h) gmail.list_inbox() calendar.today() habits.state()

⚡

Act

Synthesize via AI (Claude CLI → Gemini)

The assembled context is passed to the synthesis tier: tries Claude CLI first (your subscription), falls back to Gemini API if unavailable, falls back to a raw text summary as last resort. Never blocks.

synthesis_tier=gemini model=gemini-2.0-flash
tokens_in=2841 tokens_out=612 latency_ms=1840

👁

Observe

Deliver to Discord · write to memory

Result written to agent_runs table, appended to USER.md (episodic memory), and posted to your Discord agent channel. The next synthesis automatically includes this run as context.

insert_agent_run() append USER.md post_discord()

Scheduled jobs that deliver to Discord:

morning_briefing	7:00 AM daily · synthesis + email + calendar + tasks
evening_review	9:00 PM daily · synthesis + habit check
task_staleness	8:00 AM daily · nudge for tasks idle 48h+
relationship_decay	Mon 9:00 AM · surface fading relationships
pattern_recognition	Sun 6:00 AM · weekly deep analysis
weekly_review	Sun 6:00 PM · topic clusters + forgotten items
idea_roulette	Mon 10:00 AM · surface a forgotten old thought
memory_consolidation	2:00 AM daily · promote facts to MEMORY.md

Type any question or request in your interactive Discord channel. The ReAct loop (Reason + Act) plans which tools to call, executes them, observes the results, and replies — all in one response.

Discord conversation

ram today at 3:45 PM

research my asyncpg pool task and find me the best approach with links

🤖

2BAgent APPtoday at 3:45 PM

🤖 Researching — cycle 1/3

Found task: "Migrate legacy asyncpg connection to pool" (3 days old)

Searching for best practices…

✅ Research complete

Recommended approach: asyncpg.create_pool(min_size=2, max_size=10)

Key finding: Your existing shared _conn pattern causes "another operation in progress" errors under concurrent reconcile + job execution — confirmed in your own log from Jun 5.

Resources:
• asyncpg Pool docs — connection lifecycle, acquire/release
• pgbouncer vs asyncpg pool — when to use each
• Your own insight from 2 weeks ago: "asyncpg bottleneck on 50+ writes" aligns exactly with this

Next step: Replace _conn in services/agent/database.py with create_pool() and update all callers to use async with pool.acquire() as conn.

ReAct loop — internal trace

🧠

Reason · cycle 1

Parse request, plan tools

"User wants me to find an asyncpg pool task and research it. I'll search tasks first to get the exact text, then call web_search for implementation guidance."

list_tasks("asyncpg")

⚡

Act · cycle 1

Call list_tasks tool

list_tasks → {
"found": 1,
"task": "Migrate legacy asyncpg
connection to pool",
"age_hours": 72,
"status": "open"
}

🧠

Reason · cycle 2

Task found — now research

"I have the exact task. Now search for asyncpg connection pool best practices and cross-reference with the user's existing insights about the bottleneck issue."

web_search("asyncpg create_pool best practices") semantic_search("asyncpg bottleneck")

👁

Observe · synthesize

Merge results → final reply

Web results + user's own insight from 2 weeks ago merged. Cross-reference with agent log (Jun 5 error confirms the concurrency issue). Synthesize into actionable reply with specific file + code change.

Available tools in the interactive ReAct loop:
list_tasks · complete_task · reopen_task · research_task · add_task_note · semantic_search · list_recent · get_stats · web_search · get_calendar_event · list_calendar_events · list_emails

Full task lifecycle via Discord slash commands. Capture in any format, list open tasks, mark done, reopen, or ask the agent to research what to do next.

Open task list — /todo list

Open Tasks 4 open

Migrate asyncpg connection to pool in database.py

⚠ 3d

Review PR #47 for asyncpg pool changes

⚠ 4d

Follow up with Alice re: Q2 planning meeting agenda

⚠ 7d

Set up Ollama watchdog cron job on this machine

After /todo done asyncpg pool

Recently completed done

✓

Migrate asyncpg connection to pool in database.py

just now

Discord slash commands

/todo add Add a task. Accepts any format — "ToDo:", "todo -", "TASK:", or plain text.

/todo list [status] List tasks: open (default), done, or all.

/todo done Fuzzy-match a task by partial text and mark it complete. The agent confirms the match before acting.

/todo reopen Reopen a completed task — sets status back to open with a note.

/todo research Triggers the ReAct loop to research the task and return concrete next steps + links.

/todo note Append a context note to a task without changing its status.

/tasks Quick shortcut — same as /todo list open.

Case-insensitive capture — all map to type=task

extractor.py — supported prefixes

# All of these → metadata.type = "task"
"TODO: migrate the cron jobs"
"ToDo: migrate the cron jobs"
"todo - migrate the cron jobs"
"task: migrate the cron jobs"
"TASK: migrate the cron jobs"
"Action: migrate the cron jobs"
"Fix: the reconcile loop error"
"Review: PR #47 before Thursday"

# Also works via /todo add (no prefix needed):
"migrate the cron jobs" → task

Source	Path	What happens	Tech
Discord /capture	Bot → :8900/	Thought extracted, embedded by Ollama, stored in PostgreSQL with vector + FTS index	PythonasyncpgOllama
MCP capture tool	stdio → FastAPI	Claude Desktop calls `create_task` or `semantic_search` MCP tool → HTTP to :8900	MCP SDKDocker
APScheduler 7am	Runner → Ollama → Claude CLI → Discord	Morning briefing: fetch recent thoughts + Gmail + Calendar → synthesize via Claude/Gemini → post to agent channel	APSchedulerClaude CLI
Discord /ask	Bot → :8766/ask → ReAct loop	Interactive agent: multi-step Reason + Act cycle across up to 12 tools, reply to Discord channel	ReActDiscord.py
Gmail fetch	Runner → gmail.py → Google API	OAuth2 token refreshed, inbox fetched, body wrapped in guardrails XML before synthesis prompt	OAuth2guardrails
Search query	Client → :8900/search → DB	Query embedded by Ollama, hybrid SQL merges cosine (0.7×) + FTS ts_rank (0.3×), returns ranked results	pgvectorOllama
Memory consolidation 2am	Runner → consolidator.py → USER.md / MEMORY.md	Nightly: promote durable facts from episodic USER.md to curated MEMORY.md via AI synthesis	PythonClaude/Gemini

2BAgent in action

See every layer, step by step

Every service, mapped

Five services, one stack