PIMP MY IDEBack to dispatches
Agent memory / retrieval / operating rules

Session memory selector

Search finds the old fix. It does not make it true.

Your transcript is evidence, not policy

An old coding-agent session can contain the exact command, file, or diagnosis you need. It can also contain a workaround for a dependency that no longer exists. Retrieval solves the first problem. It does not solve the second.

Dropbox's Witchcraft repository includes a Pickbrain example that can index local sessions from several coding tools into a single SQLite-backed search system. The project says it combines semantic and keyword search, keeps the data on the device, and can return a command that resumes a matching session. We inspected the repository and its October 6 commit. We did not build or run it.

A search hit should reopen an investigation. It should not close one.

Witchcraft names XTR/WARP as its basis. The original WARP repository identifies CPU-optimized retrieval, with GPU use still recommended for index construction. SQLite's FTS5 documentation covers the other part of hybrid search: full-text queries and a rank value that can be ordered.

Memory has different jobs

Anthropic's current Claude Code documentation separates instructions that people write from notes the tool records. It says each session starts with a fresh context window. It also says instruction files and auto memory are loaded as context, not enforced configuration.

That gives us four different objects:

  1. Transcript. A record of what happened in one session.
  2. Search index. A way to find a relevant record.
  3. Project instruction. A reviewed rule that future sessions should see.
  4. Enforcement. A hook, test, sandbox, permission, or CI gate that can block the wrong action.

Do not collapse those into one bucket called memory. A command copied from a transcript is not a team rule. A team rule is not a hard control. If a correction has repeated, write the smallest durable instruction that would have prevented it. If the consequence matters, add a mechanism that can fail closed.

Use a promotion test

Before promoting a search hit, inspect its source and age. Compare it with the current tree. Then write the rule in present tense and scope it to the files or task that need it. Keep the link back to the original session or commit so another reviewer can see why the rule exists.

The bench below combines search, preview, review steps, and a copyable handoff card. It uses teaching fixtures. It does not read your history, verify your repository, or save a rule.

Interactive teaching fixture

Salvage sorter

Traditional purpose: find a prior session and reuse it. This version keeps the candidate, inspection steps, and promotion boundary on one route. The controls only build a review template.

SEARCHINSPECTCOMPAREPROMOTE
3 fixture hits. Type to filter by source, command, or rule.
Choose one candidate
Build the review template
TEMPLATE STARTED0 review steps selected

Session transcript

SOURCE TYPE / TRANSCRIPT / 41 DAYS OLD

Use npm test in packages/ui before the commit.

Risk: the command may predate a package move or script rename. Inspect the source before reuse.

Sources read / October 10, 2026 UTC

Source log

Open the evidence boundary
  1. Dropbox, Witchcraft repository, README and source read at commit cb3beef, dated October 6, 2026. The repository supports the claims about its Rust and SQLite implementation, local operation, hybrid search, Pickbrain integrations, and resume output. We inspected source but did not build, install, or run the project.
  2. Jens Lehmann Scheerer et al., XTR/WARP repository, read October 10, 2026. The repository supports the implementation basis, CPU-optimized search claim, runtime list, and GPU recommendation for index construction.
  3. SQLite, FTS5 Extension documentation, read October 10, 2026. This is the source for FTS5 full-text queries and the rank column used for ordering results.
  4. Anthropic, "How Claude remembers your project", read October 10, 2026. The page separates project instructions from auto memory, states that sessions begin with fresh context, and says those files shape behavior rather than enforce it.
  5. Hacker News discussion for Witchcraft, read October 10, 2026. This was a discovery route, not evidence for the implementation claims above.

Evidence boundary: the interactive bench uses invented fixture text to teach a review process. It searches no local files, records no evidence, writes no memory, and enforces no policy. Repository claims belong to their publishers. The operating recommendation is Pimp My IDE's.