Mempalace

MemPalace's core insight is that memory has two layers with different requirements: the content layer must be exact, and the pointer layer must be small enough to scan fast. Most memory systems conflate the two — they summarize to fit a context window and lose the exact phrasing in the process. MemPalace splits them. Drawers hold verbatim text permanently ('kept as spoken. never rewritten'). Above them, the AAAK dialect compresses each drawer into a dense symbolic address like 'D-007 @p noah ~ son.age=6 ~ dob=09-12 @l glebe-pt-rd.park @e birthday ~ party(n≈8)'. The model scans thousands of these in one pass, picks the drawers it needs, then opens them at full fidelity. That is a genuinely smart architecture, and it is why the retrieval path stays fast without paraphrasing anything. The second strong idea is entity-first organization. Wings are keyed to a real entity — a person by name, a project by codename, a domain of your life — and only inside a wing do you get rooms (days, sessions, threads), then closets (topics), then drawers. Walking a corridor unfolds the palace chronologically, room by room. Combined with the SQLite knowledge graph that stores relationships with valid-from and valid-to dates, you can ask what was true in March and what replaced it in June — 'who worked on what, when did this change, facts that were true then and may not be now.' That temporal dimension is the part most competing memory tools simply don't have. Where it will strain you: everything runs through the CLI. 'mempalace init ~/projects/myapp' scans a directory, detects people and projects from file content, and creates rooms from your folder structure; 'mempalace mine ~/chats/ --mode convos --extract general' classifies conversations into decisions, preferences, milestones, problems, and emotional context; 'mempalace search "why did we switch to GraphQL"' returns matches. After that one-time setup the docs say you shouldn't be typing commands at all — your AI calls mempalace_search for you via MCP integration, the bundled Codex plugin, or the Claude Code plugin, and hooks file new memories in the background. There is also documentation for Android/Termux via a Debian PRoot container, Gemini CLI, Cursor IDE hooks, a Remote / Team Server, and a Shared Brain, so the 'solo only' story is softer than the seed data suggests — but the core workflow is still terminal-shaped and the palace still lives in one directory on one machine until you stand up that server. The honest operational caveats: installation recommends 'uv tool install mempalace' (pip works too), Python 3.9+ and chromadb>=0.5.0 are required, and the docs carry an explicit security warning that the lookalike domain mempalace.tech is brand-squatting and runs ad-redirects and potential malware — install only from the official GitHub repository or PyPI. Compare it against cloud-synced memory assistants and you are trading GUI polish, multi-device sync, and collaboration for data sovereignty and