amplifier-tui

Cross-Session
Intelligence

When Your TUI Remembers Everything

Active Phase 4 /project ask /project search collective memory

February 2026

The question that started everything:

What if your development environment remembered everything you've ever done?

Not just the current session. Not just the last few conversations.
Everything.

The Starting Point

The most-used Amplifier interface

~750+ sessions from one developer alone

amplifier-tui evolved from a simple terminal UI to a full web-native frontend — markdown rendering, Ctrl+K command palette, collapsible sidebar, structured event cards, keyboard shortcuts, ARIA accessibility. ~102 web frontend tests.

~33,645

Lines of Code

~1,334

Test Functions

~246

Commits

The Problem

Every session is an island

🔇

Sessions don't talk

You solved the auth bug in Tuesday's session. Friday's session has no idea. You explain it again.

🔍

Finding past work is manual

"I know I did something with the caching layer last week..." Good luck scrolling through hundreds of sessions to find it.

📊

No project awareness

Sessions exist in a flat list. No grouping by project, no themes, no patterns across your work.

🧠

Knowledge evaporates

Every insight, decision, and pattern discovered in a session lives only in that session's transcript.

The Solution

A 4-phase rollout

Turn isolated sessions into a growing knowledge base — incrementally.

1

Tag Foundation

Basic tag infrastructure — sessions can be tagged, filtered, and sorted

2

Auto-Tagging

LLM reads every session and assigns tags automatically

3

Project Aggregation

Tags and sessions group into projects with counts and activity

4

Cross-Session Intelligence

/project ask and /project search — query across everything

Phase 1 — Tag Foundation

Give sessions a vocabulary

A JSON-backed TagStore that lets sessions carry metadata. Tags are normalized, filterable, and visible in the sidebar.

# Tag sessions manually
/tag auth-refactor
/tag caching performance

# Filter sessions by tag
#auth          → finds all auth-tagged sessions
#performance   → finds performance sessions

# Sort by tag groups
/sort tag       → groups sidebar by first tag

What it stores

// tags.json
{
  "sess_a3f2": ["auth", "refactor"],
  "sess_b7c1": ["caching", "perf"],
  "sess_d4e9": ["auth", "testing"]
}

Key decisions

Tags normalized (lowercase, stripped). #prefix search matches tags only. Sidebar shows up to 2 tags inline per session.

Phase 2 — Auto-Tagging

The LLM tags your sessions for you

A queue-based, non-blocking system that reads session transcripts and assigns 1–3 tags. Uses existing tag vocabulary for consistency. Respects user overrides.

Untagged Session

→

AutoTagger Queue

→

Haiku (50 tokens)

→

1–3 Tags

Smart design choices

✓ Batched: max 3 sessions per 30s cycle
✓ Uses Haiku for cost (max 50 tokens)
✓ Prefers existing tags for consistency
✓ Tracks user removals — never re-tags
✓ Graceful degradation if no API key

# The prompt (simplified)
"""
You are tagging an Amplifier session.

Existing tags (prefer these):
{tag_vocabulary}

Session project: {project_name}
Session transcript (recent):
{transcript_text}

Generate 1-3 short tags.
Output only the tags, one per line.
"""

Phase 3 — Project Aggregation

Tags become structure

Sessions and tags aggregate into projects. The sidebar shows session counts. /project gives you a bird's-eye view of all your work.

/project

Projects (4 active)

amplifier-core         42 sessions
  Tags: kernel, hooks, providers, refactor, testing
  Last active: 2h ago

amplifier-tui          38 sessions
  Tags: web-frontend, sidebar, commands, streaming, tags
  Last active: now

amplifier-foundation   27 sessions
  Tags: bundles, agents, recipes, explorer
  Last active: 1d ago

amplifier-stories      12 sessions
  Tags: decks, presentations, deploy
  Last active: 3h ago

ProjectAggregator

Stateless service. Groups sessions by directory path. Merges tag frequencies. Supports exact, case-insensitive, and partial name matching. ~145 lines of code. Zero state to manage.

Phase 4

Cross-Session
Intelligence

Sessions stop being isolated conversations.
They become collective memory.

Phase 4 — /project ask

Ask questions across all your sessions

The AI draws on your entire project history — every session, every tag, every decision — to answer questions about your work.

/project ask what's the status of the auth work?

→ Synthesizes from relevant sessions, references specific dates and session IDs

/project ask what did I do this week?

→ Weekly rollup grouped by theme, with key decisions and outcomes

/project ask what patterns are emerging?

→ Cross-session analysis of recurring themes, common approaches, evolution

Powered by Sonnet for quality synthesis. 5-minute answer cache. Builds context from project name, top 10 tags, last 15 sessions with dates and tags.

Phase 4 — Full Command Suite

Four intelligence commands

/project ask <question>

Free-form Q&A across your entire project. Ask anything — status, history, decisions, patterns. LLM references specific sessions by date and ID.

/project search <query>

Search transcript content across all sessions. Reads last 64KB per session, returns up to 20 results sorted by recency with context snippets.

/project attention

Identifies stalled work, frequent topics indicating issues, and sessions that ended without resolution. Prioritized and actionable.

/project weekly

Weekly summary grouped by theme — not chronologically. Key outcomes, decisions, and blockers surfaced automatically.

Architecture

How it all connects

TagStore

+

AutoTagger

→

Aggregator

→

Intelligence

# ProjectIntelligence builds context from:
context = {
    "project": project.name,
    "path": project.project_path,
    "sessions": project.session_count,
    "top_tags": project.top_tags[:10],
    "recent": [
        {
            "date": s.date,
            "id": s.short_id,
            "name": s.display_name,
            "tags": s.tags
        }
        for s in sessions[:15]
    ]
}

# Then asks Sonnet to synthesize
answer = await intelligence.ask(
    project, question,
    max_tokens=500
)

Concurrent Sessions

SessionHandle + registry pattern isolates per-session state. Each handle carries its own streaming callbacks and token counters. Zero cross-talk between concurrent sessions.

Answer Caching

5-minute cache keyed by MD5(project + question). Same question in the same project returns instantly. Different phrasing gets a fresh answer.

Bonus Feature

`/commit` — one command, full workflow

/commit
→ git add -A
→ generates diff summary
→ LLM writes conventional commit message
→ git commit
→ git push

/commit nopush
→ same flow, skips push

/commit info
→ dry run: branch, staged, modified, untracked

Why it matters

Runs in the active session's project directory — not the TUI launch dir. Generates conventional commit format (feat:, fix:, refactor:). Includes Amplifier co-author trailer.

Combined with /project ask, you can ask "what did I change today?" then commit it all in one flow.

The Shift

From terminal to knowledge base

Before: Pretty Terminal

✗ Sessions are disposable
✗ Flat session list
✗ Manual organization
✗ Knowledge locked in transcripts
✗ "What did I do last week?" — ¯\_(ツ)_/¯

After: Collective Memory

✓ Sessions are permanent records
✓ Organized by project and tags
✓ Auto-tagged by AI
✓ Knowledge queryable across sessions
✓ "What did I do last week?" — /project weekly

Velocity

Built in approximately one day

All four phases reportedly shipped on February 12, 2026 — in ~26 minutes of consecutive commits.

4

Phases shipped

~1,393

Lines added

~6

New modules

~26

Minutes, ~4 commits

11  47f3cde  feat: tag foundation — Phase 1
23  425b58e  feat: auto-tagging system — Phase 2        +524 lines
32  17bc18c  feat: project aggregation — Phase 3        +246 lines
37  2957e4a  feat: cross-session intelligence — Phase 4  +623 lines

Sources

Research Methodology

Data as of: February 20, 2026

Feature status: Active

Research performed:

Bundle analysis: session-discovery bundle in ramparte/amplifier-toolkit
Bundle analysis: dev-memory collection in ramparte/amplifier-collection-dev-memory
Feature inventory from bundle.md manifests and README files
Command documentation from session-discovery-awareness.md

Gaps: LOC counts, test counts, commit counts, and session counts from existing deck could not be independently verified in this research pass. Velocity numbers (~26 minutes, ~4 commits) not independently confirmed.

Repository: ramparte/amplifier-toolkit (session-discovery), ramparte/amplifier-collection-dev-memory (dev-memory)

Try It

Your sessions are already
a knowledge base.

You just need to ask.

/project ask

Query your history

/project search

Find past sessions

/project weekly

Your week, summarized

amplifier-tui — ~750+ sessions and counting

Cross-SessionIntelligence

What if your development environment remembered everything you've ever done?

The most-used Amplifier interface

Every session is an island

A 4-phase rollout

Give sessions a vocabulary

The LLM tags your sessions for you

Tags become structure

Cross-SessionIntelligence

Ask questions across all your sessions

Four intelligence commands

How it all connects

/commit — one command, full workflow

From terminal to knowledge base

Built in approximately one day

Research Methodology

Your sessions are alreadya knowledge base.

Cross-Session
Intelligence

Cross-Session
Intelligence

`/commit` — one command, full workflow

Your sessions are already
a knowledge base.