Memory & knowledge

Session knowledge base

Turn months of agent sessions into a searchable library with cards and an index.

Paid $9 list, $7 until 2026-12-31 · one skill, yours for good. The outline, the steps and the file open after purchase.

Background

Turn the sessions scattered across the DSH session store, the Codex home and DeepSeek export packages into queryable cards + index + project pages + topic pages + lessons/decisions ledgers. A one-command job is not retention; being rerunnable is.

Why it is built this way

Every step exists so the next rerun is cheap. extract.py turns sessions into JSON. make_batches.py cuts at 130k characters and 10 sessions, which is what keeps one subagent's batch inside its context. render_cards.py turns card JSON into 10-会话卡 pages, and normalize, index and catalog build the pages retrieval actually uses. check.py closes the loop by comparing formal session count against card count, with 0 missing fields, 0 plaintext keys and 0 dead links. PASS is the only finish.

Card building is delegated because the unit of work is a batch, and concurrency sits at 6 after a run that started 65 at once failed the whole batch with active child limit: 8. Curation follows the same reasoning: a lessons page with 700+ entries is never read, so it keeps 2 to 10 per group and the full set goes to *_full.md. Extraction waits for 15 quiet minutes because a session still being written yields a truncated card.

Where it bites

• Too many card subagents at once Several batches start together and dozens of card-building subagents get dispatched. The whole batch fails with active child limit: 8 and cards already built go with it. The first hard rule caps concurrency at 6, since 8 was measured as the ceiling.
• Incremental rerun in the wrong order Rerunning with make_batches before render_cards means extract has already cleared the card field in sessions/*.json, so every session that has a card counts as missing and a long list of false pending items buries the real ones. The order is extract, render_cards, make_batches.
• Plaintext keys ride back into the cards Redaction covers the summary text only, while sessions/*.json keeps the raw arrays, so each re-extraction carries keys starting sk- or ark- back into the cards. The fix is a JSON rescan after every extract, rewritten to REDACTED, and check.py counts plaintext keys.
• A new batch reuses the old prefix Sessions added later get batched without --prefix, so the new b01 overwrites the old batch's card JSON and those cards disappear from disk. check.py then reports a mismatch between session count and card count. The second hard rule asks for make_batches.py --prefix w --tag dsweb.

When to use it

Use when the machine's multi-platform AI sessions (DSH / Codex / DeepSeek web export) need to be organised into a searchable knowledge base, or when an existing knowledge base needs an incremental rerun; also for requests like 'can these sessions be captured for reuse' and 'how do I top up a knowledge base that has gone stale'.