The best-benchmarked open-source AI memory system. And it's free.
Created 2026-04-05 and already at v3.3.5 as of 2026-05-10 — a 3.x release reached in ~35 days is anomalous and almost certainly indicates either a repo migration/rename or version numbers that don't map to conventional semver milestones. The scam-alert caution block, the HISTORY.md "corrections and public notices," and the impostor-domain warning collectively suggest a project with a prior public history that predates this GitHub repo. Taking the artifacts at face value: reproducible benchmarks committed to the repo, a changelog, a contributing guide, 29 documented MCP tools, a full external docs site, and a Discord community all point to a project well past alpha. Daily commit cadence and a patch release the day before appraisal confirm someone is actively steering it. The version ambiguity prevents a 5.
README is production-grade: clear problem statement, quickstart with copy-pasteable commands, explicit benchmark reproduction steps, benchmark caveats (teaching-to-test risk flagged honestly in the rerank section), and a full external docs site segmented into guide, CLI reference, Python API, concept pages, and MCP tool reference. BENCHMARKS.md, CHANGELOG.md, CONTRIBUTING.md, and HISTORY.md are all present and linked. The two-tier approach — README for orientation, docs site for depth — is exactly right. No gaps visible at this detail level.
Positive signals: pluggable backend with an explicit interface in mempalace/backends/base.py, SQLite-backed knowledge graph (appropriate for local-first tooling), ChromaDB as the default vector store (well-maintained upstream), Python 3.9+ baseline covering wide runtime compatibility, and uv as the recommended installer (signals awareness of modern Python packaging hygiene). Dependency manifest was unavailable so transitive dep hygiene can't be verified. No direct visibility into test suite or CI configuration from the provided data, though the benchmarks being committed and reproducible is a meaningful proxy for automated verification discipline. Architecture descriptions suggest clean separation of concerns.
Last commit 2026-05-11 — one day before appraisal. Latest release v3.3.5 published 2026-05-10. 226 open issues indicate an engaged community and active triage surface. Discord linked from README. The banner warning about Claude Code session expiry referencing a specific discussions thread (#1388) shows the maintainers are actively translating community pain into documentation. No signs of abandonment at any signal layer.
52,054 stars and 6,856 forks are very large absolute numbers. The growth rate implied by the April 5 creation date (~1,400 stars/day average) is extraordinary and should be noted — it is either a viral debut, a migration carrying prior star history, or inflated. The presence of impostor domains sophisticated enough to warrant a repo-level caution block and a dedicated HISTORY.md is a credible secondary signal of real adoption. Forks at 13% of stars is a healthy ratio suggesting active derivative use rather than passive starring.
Overall: 4.6/5
Category: Personal AI Memory Known alternatives in vault: garrytan--gbrain (4.3/5, current top pick), NorthwoodsSentinel--loam (2.2/5), UnluckyMycologist68--palimpsest (1.1/5) Differentiation: MemPalace's core differentiator is verbatim storage — it explicitly refuses to summarize or paraphrase, preserving original content in a hierarchical palace structure (wings → rooms → drawers) that enables scoped retrieval rather than flat-corpus search. The benchmark transparency is exceptional: numbers are reproducible from the repo, held-out vs. tuned splits are honestly reported, and the README explicitly declines to publish misleading cross-system comparisons. The 29-tool MCP server, temporal knowledge graph with validity windows, agent diary/wing isolation, and auto-save hooks for Claude Code sessions are all features beyond what the current vault alternatives demonstrate. gbrain, the current top pick, likely competes on simplicity and tighter gbrain-evals integration but lacks MemPalace's retrieval depth, MCP surface area, and benchmark rigor. Gap or crowd: Category has adequate coverage (3 repos) but is dominated by weaker entries. MemPalace would become the new top pick by a significant margin, displacing garrytan--gbrain. The category remains a healthy addition to keep — MemPalace is not redundant, it upgrades coverage quality.
Score: 5/5
Harvestable: The palace metaphor (wings/rooms/drawers) as a structured knowledge-organization primitive; the hybrid retrieval pipeline (semantic + keyword boosting + temporal-proximity + preference-pattern extraction) as an algorithm pattern; the mempalace wake-up session-context-loading pattern; the sweep command for per-message verbatim drawer creation; the temporal entity-relationship graph with validity windows; the MCP tool architecture for 29 discrete operations across palace reads/writes, agent diaries, and knowledge-graph CRUD.
Integration path: Near-zero friction for the core path: uv tool install mempalace → mempalace init → wire the two Claude Code auto-save hooks (periodic + pre-compression) → configure MCP server in Claude Code settings. mempalace mine ingests existing project files and conversation history immediately. mempalace wake-up slots directly into a session-start skill. The 29 MCP tools are immediately callable from any MCP-compatible orchestration layer without additional wrapping.
Overlap with existing: garrytan--gbrain is the primary overlap — both serve AI memory retrieval for the same PAI use case. NorthwoodsSentinel--loam overlaps on structured knowledge storage. garrytan--gbrain-evals overlaps on retrieval benchmarking methodology. MemPalace is a strict superset of what those three repos provide for this use case; it does not introduce category-level redundancy against any other vault repo.
Adoption cost: Trivial. The CLI-first design, uv tool install packaging, and explicit Claude Code hook documentation mean integration is a configuration task, not a development task. The MCP server requires only adding a server entry; no custom glue code is needed to expose the 29 tools.
Three flags worth tracking: (1) The version-to-age ratio (v3.3.5 in 35 days) and the star-to-age ratio (~1,400/day) are both statistical outliers. The most benign explanation is a repo migration with carried history; the scam-alert caution block and HISTORY.md "corrections and public notices" section support a project with a complicated prior public identity. This does not impair current utility but is worth confirming against HISTORY.md before deep integration. (2) The README's "Important" banner about Claude Code session expiry is an explicit hook-wiring CTA — it is both useful and a signal the maintainers understand the Claude Code integration use case intimately, which bodes well for hook reliability. (3) The benchmark section is one of the most honest in the AI memory space: the 98.4% held-out hybrid figure is correctly distinguished from the tuned figure, the rerank ≥99% is explicitly flagged as potentially teaching-to-test, and cross-system comparisons are explicitly refused. This epistemic discipline is a strong positive quality signal for a project making bold benchmark claims.