Current stage: Windows desktop development build. Updated 10 October 2026.
A coding conversation can explain why a database was chosen, which approach was rejected, or what should happen next. The code records the result. The conversation often holds the reasoning.
That is the problem behind Mind Palace: keep coding conversations available locally, then find the original passage when the question comes back.
The question I use in the demo is deliberately small: “Why did we choose SQLite?” The answer is not generated. The app finds a stored message explaining that SQLite keeps storage on the device and needs no separate database server. A source button opens the saved conversation so the reader can check it.
The demo uses fictional messages in a real native app. It demonstrates the storage and retrieval path, not a live model conversation.
Mind Palace uses Angular for the interface, Tauri for the desktop boundary and bundled Python for capture, storage and search.
Angular handles the setup checklist, session Library and Ask screen. Tauri exposes narrow, validated native commands and owns the worker processes. Python preserves source revisions, imports captured sessions into a vault and searches those sources.
The flow is:
The current capture, storage and search path needs no API key or model call. The application enables no developer analytics or hidden cloud sharing.
An edited summary should not replace the conversation that produced it. Mind Palace preserves selected source revisions as immutable, content-addressed files. New transcript prefixes can update the canonical session without overwriting older source revisions.
One provider/session identity maps to one canonical Session. Reimporting a capture should not create another copy of the same conversation. Ambiguous or divergent revisions remain visible failures instead of being silently treated as the newest source.
This distinction also matters for decisions. A suggestion is not an approval, and reviewing a summary does not confirm its proposed decisions. The repository keeps those boundaries explicit rather than treating every assistant statement as agreed memory.
Ask performs bounded keyword matching over verified stored conversations. It returns an exact passage with a source reference, or reports that it found no matching evidence. It does not synthesize explanations or provide semantic search today.
That narrower behavior makes the output inspectable. It also has limits: different wording can miss a relevant passage, and a matching passage still needs human interpretation. “I found these words” is not proof that a decision remains current.
Recorded native checks cover closed-app capture, restart import, exact-source recovery, local Ask and source opening using isolated synthetic data. A complete 20 MiB source passed capture/import/read with matching SHA-256 in an earlier verification run.
Those results do not make this a finished release. Packaging and broader reliability work remain open, including a recorded 18-of-20 event-delivery discrepancy in a separate controlled Codex test. Claude Code end-to-end support and generated model answers are not current product claims.
Mind Palace on GitHub includes a browser sample, source, screenshots and setup instructions. The browser sample is fictional, resets on reload and cannot connect a desktop vault. The native workflow requires a prepared Windows build; the repository's development guide explains that distinction.
Start with the user guide or read the verification record.
If you use coding agents, which past decision is hardest to find again? Describe the problem in a repository issue without sharing private conversation content.