Every Session
Was An Island

Cross-session memory, synthesized from local data

The Problem

Every session was an island, with no memory of the last

amplifier-tui piles up work as many small, focused sessions. But today each one is an island — Friday's session has no memory of what Tuesday solved.

And that isolation isn't a small problem — it's happening at scale.

# Today each session is an island.
# (DESIGN.md:5, CONTEXT.md:11)

session A  →  ??? 
session B  →  ???
session C  →  ???

# nothing connects them

Proof — At Scale

3,324 transcripts across 375 projects sit unsynthesized

The island problem is real and large. On one machine, the local Amplifier session store already holds thousands of transcripts — none of them talking to each other.

The data is clearly there. So why can't a session remember?

3,324

transcript.jsonl files

375

project directories

The Real Gap

The data to connect them already exists locally

Transcripts, metadata, tags, and scanner state all live on disk — but nothing synthesizes them into project-level understanding.

The gap isn't collection. It's synthesis. That's exactly what shipped next.

✓

Transcripts

Already on disk

✓

Tags & metadata

Already on disk

✓

Scanner state

Already on disk

×

Synthesis

The missing piece

The Fix — One Merge

One squash merge shipped a four-phase Project Intelligence system

PR #2 (commit eb8ee2e, merged 2026-02-12) landed the whole answer at once: an ordered set of phases that turn isolated sessions into collective memory.

The first design choice is what makes it cheap: no new datastore.

1

Tag Foundation

Reuses the existing TagStore

2

Auto-Tagging

Background, cheap model

3

Project Aggregation

Stateless aggregator

4

Cross-Session Intelligence

Ask, weekly, attention, search

How — No New Plumbing

No new datastore — the working directory is the project

A stateless ProjectAggregator runs on demand, treating the working directory as the project boundary. It reads the tags and sessions already loaded — a view over existing data, not a new store.

With the data aggregated, one engine turns it into answers.

≡

Stateless aggregator

Runs on demand from data already loaded

⌂

Project = working directory

No new "project" entity

▤

Reuses TagStore

No schema change needed

How — The Synthesis

A sonnet-class engine synthesizes answers, all on-machine

ProjectIntelligence uses claude-sonnet-4-20250514 for quality synthesis; background auto-tagging uses the cheaper claude-haiku-4-20250506. All data stays local — no new data leaves the machine.

Which means you can finally just ask.

S

Synthesis: sonnet

claude-sonnet-4-20250514

H

Auto-tagging: haiku

claude-haiku-4-20250506

🔒

All data stays local

Nothing leaves the machine

The Payoff

Now you just ask, in plain English

Answered from your own local sessions and cached for 5 minutes — with dedicated weekly, attention, focus, and search subcommands beside it.

Isolated sessions became queryable collective memory.

/project ask what did I do this week?

# dedicated subcommands
/project weekly
/project attention
/project focus
/project search <query>

The Takeaway

1,790 lines turned islands into memory

Five files, built by one developer, connected the sessions — because the data was there all along. The value came from synthesis, not collection.

Wire up the local data you already have, and isolation dissolves.

1,790

lines across 5 files

1

squash merge (PR #2)

Sources

Research Methodology

Data as of: verified against ramparte/amplifier-tui (git@github.com:ramparte/amplifier-tui.git)

Feature status: SHIPPED · Phases 1–4

Research performed — commands run in a fresh session:

git show -s --format='%an | %ci | %s' eb8ee2e · gh pr view 2 --json number,title,author,mergedAt,state — PR #2 squash merge eb8ee2e, merged 2026-02-13T05:21:03Z (2026-02-12 PT)
grep -n '^## Phase' project_design/IMPLEMENTATION.md — four-phase design (plus Phase 5 Web)
grep -rn 'island' project_design/ — "Today each session is an island" (DESIGN.md:5, CONTEXT.md:11)
find ~/.amplifier/projects/*/sessions -maxdepth 2 -name transcript.jsonl | wc -l · ls -d ~/.amplifier/projects/*/ | wc -l — 3,324 transcripts across 375 project dirs (dynamic local metric)
grep -n 'model' .../project_intelligence.py .../auto_tagger.py — sonnet claude-sonnet-4-20250514 / haiku claude-haiku-4-20250506
grep -n '_CACHE_TTL_SECONDS' .../project_intelligence.py — 300s (5-minute) answer cache
wc -l across auto_tagger.py, project_aggregator.py, project_intelligence.py, project_search.py, project_cmds.py — 1,790 lines total
git log --format='%an' | sort | uniq -c | sort -rn — 283 of 284 commits by Sam Schillace (GitHub login ramparte); span 2026-02-05 to 2026-04-23

Gaps: Transcript/project counts are dynamic local-machine metrics (drift upward as background scanning adds sessions). No dedicated unit tests for the aggregator/intelligence/auto-tagger/search modules were found; the feature is not claimed to be unit-test-covered. Phase 5 (Web) is designed/directional, not independently verified.

Primary contributor: Sam Schillace (GitHub login: ramparte) — 283 of 284 commits.

Every SessionWas An Island