MemPalace is an open-source AI memory system designed to give LLM agents persistent, structured, and queryable memory across sessions. It organizes codebases and documents into a hierarchical "palace" structure — Wings (projects), Rooms (directories), and Drawers (files) — powered by ChromaDB under the hood for fast similarity search. The project was created to solve the context-window exhaustion problem that plagues long-running AI coding agents: instead of losing history when context fills up, MemPalace summarizes and compresses older content into retrievable memory nodes, enabling agents to recall relevant context on demand.

• Hierarchical Palace Architecture — Wings → Rooms → Drawers provides a familiar spatial metaphor for organizing knowledge. Agents can navigate "rooms" just like navigating a file system, and the system supports MCP (Model Context Protocol) tools so AI agents can query, add, and update their own memory during live sessions.
• Adaptive Memory Compaction — The system automatically summarizes and compresses older conversation history and code chunks using LLM-based compaction. Files are mined in batches, chunks are deduplicated by mtime, and the HNSW index in ChromaDB handles retrieval with configurable similarity thresholds.
• Production-Ready MCP Server — MemPalace ships with a full MCP server implementation with search, status, taxonomy, and CRUD tools. A recent PR (#562) added batch ChromaDB writes, concurrent mining, a purge command, and junk file filtering (minified JS/CSS, bundles, source maps) — making it suitable for indexing large real-world codebases.

A detailed community audit flagged major discrepancies between what MemPalace's README advertised and what the code actually did. One commenter (akarnokd) noted: "It's wild how a project can gain thousands of stars purely on hype and 'Readme-Driven Development' while the actual codebase is basically regex and unverified claims." Another (dial481) pointed out that on LongMemEval, the project's own retrieval benchmark showed a 96.6% → 84.2% AAAK drop — directly contradicting the "30x lossless compression" claim. The maintainer acknowledged the issues and multiple follow-up PRs have since addressed the technical gaps.

"I was confused because they say no LLM calls like ChromaDB isn't using embeddings lol." — mattneel, GitHub Issue #27

A major PR consolidated performance and reliability improvements. The most impactful fix was cache invalidation — there was a race condition between tool_add_drawer/tool_delete_drawer and tool_update_drawer where stale metadata was served in concurrent setups. The reviewer (web3guru888) called it "the most impactful of the three fixes — invalidating _metadata_cache after every write means status and taxonomy are always consistent." The PR also introduced junk file filtering (skipping minified JS/CSS, bundles, source maps, lockfiles > 500KB) and a new mempalace purge --wing X --room Y CLI command.

"Tested this branch against our 14,902-drawer palace — the MCP server side lands cleanly, but the CLI still has the old bug." — psaghelyi, Issue #562

A critical bug report documented three related failure modes: HNSW graph corruption causing 1.5TB+ sparse link_lists.bin files, PreCompact deadlock on macOS M4 Pro, and mine fan-out runaway. The fix addressed multiprocessing get_context("fork") unavailability on Windows CI and ensured hook_precompact runs transcript mining synchronously rather than in a potentially deadlocked thread.

"Thanks for the thorough review — all six comments were actionable. Pushed follow-up commit." — felipetruman, Issue #976

MemPalace is one of the most actively discussed open-source AI memory projects, with over 500 open issues and a rapidly evolving codebase. It stands out for its MCP server integration — making it agent-native rather than just a retrieval backend — and its willingness to engage with hard technical critiques head-on. The project has matured significantly since its April 2026 launch, with major reliability fixes landing in recent weeks around concurrent writes, HNSW corruption, and Windows compatibility. If you're building AI coding agents that need persistent context across sessions, MemPalace is worth evaluating — but keep an eye on the issue tracker for the latest fixes.

@MemPalace/mempalace · ⭐ 51,777 Stars · View on GitHub