The best-benchmarked open-source AI memory system. And it's free.
| Rating | Summary | |
|---|---|---|
| Quality | solid (18/24) | Well-documented and explosively adopted library with solid release cadence; marked down for Python-only stack, no visible CI badge, and issue backlog above triage threshold. |
| PAI Relevance | integrate (0.75) | Fills PAI's most significant architectural gap — semantic vector retrieval over conversation history — with harvestable patterns (hierarchical memory, temporal KG, verbatim verbatim vs. summary tradeoff) that directly apply to PAI's memory subsystem. |
18/24 — maintained / well-documented / solid
Failed:
Passed:
Failed:
Passed:
Failed:
uv sync --extra dev implies dev extras but no test runner or test section is referenced.Passed:
uv sync --extra dev and pip install mempalace[dev], confirming pyproject.toml exists.| Dimension | Score | Assessment |
|---|---|---|
| Harvest Value | 2 | The hierarchical wings/rooms/drawers memory structure, verbatim-over-summarization philosophy, and temporal entity-relationship graph with validity windows are all novel relative to PAI's flat WORK/LEARNING/KNOWLEDGE file store and are directly applicable to PAI's memory subsystem architecture. |
| Integration Readiness | 1 | Python-only with no TypeScript bindings; PAI could subprocess the mempalace CLI (mine, search, wake-up) from a PAI skill, and 29 MCP tools offer a second integration path, but both require moderate adapter code given PAI's TypeScript/Bun-only rule. |
| Overlap Risk | 1 | Partial overlap with the ContextSearch skill (session recovery search) and the Knowledge skill (persistent knowledge graph); MemPalace's semantic vector retrieval layer has no equivalent in PAI's current capability manifest. |
| Gap Fill | 2 | PAI has no vector-indexed semantic retrieval over conversation history — its memory layer is purely file-based markdown/YAML. MemPalace's core function (96.6% R@5 verbatim retrieval at zero API cost) directly addresses this gap. |
Composite: 0.75
Every Claude Code session currently starts blank, forcing repeated re-explanation of project decisions and conventions across Conservancy, Aphoria, and the Capture pipeline: Run pip install mempalace && mempalace mcp start, add the MCP server endpoint to ~/.claude/mcp_config.json, and add a one-line CLAUDE.md note to each active project repo instructing the agent to write decisions, dead ends, and agreed conventions to MemPalace before closing. All 29 MCP tools become available immediately with no API key — agent sessions accumulate a searchable cross-project memory that persists across restarts.
The Capture-to-Knowledge pipeline has no fast retrieval layer for validated knowledge, meaning every new capture has no way to check whether equivalent content already passed validation: Wire MemPalace as the post-validation write target — after each clean-room Haiku validation pass, call mp.store(content, room="validated-captures") via the Python API, then query mp.search(new_capture) at intake to surface near-duplicate prior captures before committing validation effort. Because MemPalace hits 96.6% R@5 with zero LLM at inference time, this deduplication check adds negligible latency and no token cost to the pipeline's hot path.
fab currently recommends Fabric patterns with no memory of which recommendations were accepted or discarded, meaning the same poor matches can recur across invocations: After each fab run, append the content fingerprint, the recommended pattern, and the accept/reject outcome into a dedicated fab-decisions wing via mp.store({"fingerprint": ..., "pattern": ..., "outcome": ...}, room="outcomes"). On subsequent runs, add a pre-recommendation step that queries MemPalace for the closest prior fingerprint and filters out patterns previously marked rejected for that content type — turning the tool's recommendation history into a local, zero-cost prior that improves with use.
Category: Personal AI & Knowledge
In this category: tinyhumansai--openhuman (excellent, 21/24, skip) — previously appraised; MemPalace--mempalace is first memory-system-specific entry at the full appraisal level.
Standing: Where openhuman is a general personal AI superintelligence layer, MemPalace is narrower and deeper — purpose-built memory infrastructure with reproducible benchmark numbers, placing it as the most rigorously validated memory system in the category.
Density: 9/10 — Available: full README (8KB), stars, forks, issue count, release tag and date, creation date, last commit date, language, license, topics, description, landscape context with overlap clusters, prior appraisal stubs. Missing: dependency manifest (pyproject.toml contents not retrieved), CI configuration, contributor count, test suite details.
The "poor (4.6/24)" placeholder in the rolling summary reflects an early uncategorized stub before this full appraisal; the actual assessment is solid (18/24). The security warning about impostor domains in the README opening is a notable signal: the project attracted impersonators within weeks of launch, consistent with the 53k-star adoption velocity. The benchmarking infrastructure (reproducible from the repo, per-question result files committed, explicit methodology documentation) is unusually rigorous for a project seven weeks old and sets MemPalace apart from other memory systems in the landscape that rely on self-reported or non-reproducible numbers. The 249 open issues on a 52-day-old repo is a flag to watch — rapid growth without proportional maintainer bandwidth creates triage debt.