One session library across three machines

2026-10-11

I keep long chat threads with coding agents alive, because a new chat starts with no context. That is a bad habit. I built a library so a new chat can look up what an old one decided.

The problem

Context sits in four places: memory files, worklogs, docs/ files, and handoff files. Memory is tied to the folder of a session, so a new git worktree starts empty. Agents also re-read raw transcripts to recover context. A scan on 2026-10-10 found 507 such reads in 112 sessions. That is about 225k tokens of results.

The design

• One private git repo holds the library. It syncs to all three machines.
• One markdown entry per topic, in entries/<topic>.md.
• Each entry has fixed parts: Answer, Decided, Blocker, Superseded, Measured, Open, Pointers, and the session ID.
• A generated INDEX.md lists one line per entry.
• A capture step runs inside an existing session reviewer. It runs every 15 turns and at session end. It adds no new model calls.

The lookup

A library skill tells the agent the order:

1. Read INDEX.md.
2. Read the entries that match the topic.
3. Run lib search "<keywords>", at most three times.
4. Open the full transcript of a cited session.

After three empty searches, the agent says there is no prior work. The lib CLI also has lib index and lib build-index. The search is a per-machine SQLite FTS5 keyword index over condensed, redacted transcripts. It has no MCP server and no embeddings.

The test

I tested 24 sessions and 12 questions with Haiku, one run per setup. The score is correct answers out of 12.

• Raw transcripts read in full: 12.
• Entries plus keyword search: 10.5.
• Entries plus hybrid search: 10.5, at a higher cost.
• Keyword search only: 9.5.

One run per cell is a small sample. The misses came from capture quality, not from retrieval.

What is still open

• Capture works only after a session reaches 15 turns or ends.
• Security-review sessions still show up in search results.
• A negative answer about the state of code costs about 100k tokens, because the agent then checks the code.
• Old transcripts hold plaintext logins. I strip logins only in entries. I must rotate those logins myself.