Skip to content
Agents tracked: 501 Releases (7d): 777 Agent status: no reported issuesSince last limit reset: Claude 10d · Codex 8d For AI agents: Markdown Updated Nov 21, 2026

MemPalace changelog: what's new each month

Every stable MemPalace release summarised by month: the highlights, new features, improvements, fixes and anything you need to act on. 8 months covered; the current month updates daily.

November 2026 · so far

1 release: 3.12.0

November brought significant infrastructure updates to MemPalace, including Python 3.10+ requirements, a migration to SQLite for hallways storage, and new security controls for HTTP hubs and LLM integrations.

Action needed
  • Python 3.9 no longer supported; upgrade to Python 3.10 or newer required.
  • PQL search filters now require uppercase keywords (FIND, IN); use FIND auth IN backend instead of lowercase variants.
  • Environment variables renamed: LLM_ENDPOINT, LLM_MODEL, and LLM_KEY must be prefixed with MEMPALACE_.
  • External LLM endpoints now require explicit user consent.

Highlights

  • Python 3.9 support dropped; 3.10 or newer now required.
  • Hallways now stored in SQLite instead of JSON, with automatic migration on first write.
  • PQL search syntax changed to uppercase keywords for filter operations.
  • Local HTTP hubs can now run without authentication tokens for loopback access.
  • Environment variables for LLM configuration renamed with MEMPALACE_ prefix and require explicit consent for external endpoints.

New

  • Explicit UTC offset support for new timestamps.
  • Browser-origin, Host, and proxy-header validation checks for local HTTP hubs.
  • Transactional import of hallways.json into hallways.sqlite3 with archived retention.
  • Conversation mining now preserves user turns exactly as written.

Improved

  • Mine re-reads each source it covers to apply parsing and byte-preservation fixes.
  • Drawers with unchanged text reuse their stored vectors instead of regenerating them.
  • Legacy naive timestamps handled as local time, preserving original interpretation.
  • Spell correction disabled by default in conversation mining.

Fixed

  • Normalization now correctly handles parsing and byte-preservation in source text.

October 2026

1 release: 3.11.0

October brought embedding model flexibility with new options like EmbeddingGemma 2, native media retrieval capabilities, and stronger safeguards to prevent palace corruption from model mismatches. Existing palaces remain unaffected unless you explicitly opt into new features.

Action needed
  • EntityRegistry.research() and confirm_research() methods removed (unused Wikipedia lookup feature).

Highlights

  • EmbeddingGemma 2 is now available as an opt-in embedder, with multimodal support for images, audio and video.
  • Palaces now record their embedding model and refuse to open with a different one, preventing silent corruption.
  • Native media retrieval embeds images, audio and video files into palaces for typed local search.
  • New audit command scores palace organization across five structural layers with detailed findings and CI integration.

New

  • EmbeddingGemma 2 embedder option with optional multimodal support for media files.
  • Native image, audio and video retrieval through built-in media source adapter.
  • Mempalace audit command to evaluate palace organization with layered analysis.
  • JSON output mode for audit results.
  • Fail-under threshold option for audit command in automated workflows.

Improved

  • First conversation mine after upgrading re-mines each covered conversation file once for freshness.
  • Palace opening now validates embedding model match across CLI, MCP server, miners and repair operations.

Fixed

  • Palaces no longer fail silently when opened with a mismatched embedding model.

September 2026

1 release: 3.10.0

September brought architectural changes to how MemPalace handles configuration and collections, along with performance improvements to vector operations and search.

Action needed
  • mempalace rules --agent option removed; use --host, --harness, and --project parameters and re-render any installed shared-brain blocks.
  • getcollection() now raises CollectionNameMismatchError for any collection name other than configured drawers collection or mempalaceclosets.

Highlights

  • Configuration and palace data now use XDG Base Directory specification for new installs.
  • Read-only search operations no longer block on palace writer locks, improving concurrent performance.
  • Native Rust vector acceleration now available as an optional feature for faster exact-vector operations.
  • Event listings now return newest events first by default, matching typical usage patterns.

New

  • Optional native Rust exact-vector acceleration module (rustexact).
  • Standalone native vector CLI tool and optional Python wheels.
  • NumPy top-k optimization for vector operations.

Improved

  • Event listing API defaults to newest-first ordering for better user experience.
  • Read-only search operations no longer contend with palace writer leases.
  • Collection name validation now prevents invalid getcollection() calls with CollectionNameMismatchError.

Fixed

  • mempalace rules command no longer accepts invalid --agent parameter; users must pass --host, --harness and --project instead.

August 2026

4 releases: 3.7.0 → 3.9.0

August brought major performance improvements and multi-agent coordination features to MemPalace. The month focused on scaling search and memory operations while adding foundational support for coordinated agent workflows.

Highlights

  • Agent logstream coordination layer enables durable task handoffs and synchronized work across multiple agents.
  • Search performance dramatically improved: metadata reads now skip vector indices, reducing latency from seconds to milliseconds on large palaces.
  • OpenAI-compatible embeddings allow use of any standard embeddings endpoint including self-hosted models.
  • Hub-proxied sessions are now lightweight, importing only forwarding logic until fallback is needed.
  • Search results enrichment is faster and clearer with indexed source lookups and concurrent Hub reads.

New

  • Logstream coordination with append-only event layer for durable task packets and agent handoffs.
  • OpenAI-compatible embeddings model option to use any /v1/embeddings endpoint.
  • Search date window filtering via since/before parameters for time-bounded queries.
  • Hermes memory provider that files conversation turns with matching metadata.
  • mempalace logstream watch command for durable event monitoring with cursor advancement and type filtering.
  • Skill-first guided setup for bootstrapping private or shared installations with fleet identity assignment.
  • Private-palace search benchmark for evaluating performance without exposing queries or content.
  • Logstream topic routing and reverse pagination for focused coordination.

Improved

  • Chroma metadata reads skip vector index lookups, reducing peak memory and latency dramatically.
  • SQLite exact reads no longer scan entire palace; listdrawers and search queries execute in milliseconds.
  • Long-running Chroma servers maintain flat memory usage instead of reloading indices on own writes.
  • EmbeddingGemma groups documents by size before batching to reduce padded-token overhead.
  • Hub-proxied MCP sessions are thin, lazily loading full server only if local fallback needed.
  • Write-routing now coordinates through daemon for serialized mutations in multi-session setups.
  • Release awareness is explicit and opt-in, allowing agents to prepare upgrade plans without blocking.
  • Configuration handling is safer: unreadable files preserved, writes atomic, symlinks intact.

Fixed

  • Ingest commands no longer hang when encountering named pipes in mined directories.
  • Project re-mine now tracks chunktotal correctly to detect partial or interrupted files.
  • Conversation mine completeness properly stamps drawers and cleans partial data on upsert failure.
  • Stale Chroma database cache cleared on palace reconnect after peer or rebuild changes.
  • MCP releases writer lease on SIGTERM/SIGHUP signals, fixing SSH disconnect hangs.
  • EmbeddingGemma auto-selection avoids silent NaN or zero vector storage on Apple Silicon.

July 2026

1 release: 3.6.0

July brought major infrastructure and storage improvements to MemPalace, including a new team server mode and alternative database backend, alongside enhanced knowledge extraction and temporal awareness for conversation management.

Highlights

  • New secure remote server mode lets teams run MemPalace as a shared HTTP service with built-in authentication.
  • Milvus storage backend option supports multiple deployment models with advanced vector and lexical search.
  • Atomic fact replacement preserves knowledge graph integrity when superseding information.
  • Conversation timestamps are now preserved throughout the system for better temporal queries and relevance.
  • Mining extracts structural elements like code symbols and URLs into the associative graph.

New

  • Turnkey secure team server accessible over HTTP with bearer token authentication.
  • Milvus Lite, self-hosted Milvus, and Zilliz Cloud storage backend support with COSINE and BM25 search.
  • supersede() function for atomic knowledge-graph fact replacement with shared timestamps.
  • Per-project mining exclusions via excludepatterns in mempalace.yaml.
  • LaTeX source file support for .tex and .bib prose mining.
  • Structural entity extraction recording code symbols, URLs, paths, and qualified identifiers.

Improved

  • Conversation chronology preserved with authored-at timestamps in drawers and search results.
  • Hybrid search scoring prefers recently authored content when exact scores tie.
  • Mining exclusions work without changing --limit semantics.
  • Shared daemon write-routing policy provides consistent routing for write-capable clients.
  • Mined sessions now populate the associative graph for better knowledge connections.

June 2026

4 releases: 3.3.6 → 3.5.0

June brought major enhancements to MemPalace's knowledge graph and IDE integration, including multilingual support by default, pluggable vector backends, and native Cursor and Google Antigravity IDE plugins. The month also strengthened connection tracking between entities and added new transcript parsers for AI tools.

Highlights

  • Office-document mining now supports PDFs, Word, PowerPoint, Excel, RTF, and EPUB files alongside existing formats.
  • Cursor IDE and Google Antigravity IDE plugins with auto-registered slash commands and background memory recall.
  • Hallways and tunnels: automatic connection tracking between entities within and across wings, with Hebbian learning to strengthen repeated connections.
  • Multilingual embeddings now default for new installs, replacing English-only models.
  • Pluggable vector backends (Qdrant, pgvector, sqliteexact) alongside ChromaDB.

New

  • Virtual line numbers and surgical closet pointers for exact line range citation in drawers.
  • Cross-wing tunnels automatically promote shared entities across multiple wings for unified queries.
  • Living-memory dynamics with Hebbian potentiation and Ebbinghaus decay on hallways and tunnels.
  • API-tool transcripts auto-route to wingapi (Claude Code, ChatGPT, Slack-bot exports, OpenAI-shape JSON).
  • Opt-in local daemon for queued writes to serialize palace writes and prevent race conditions.
  • Optional HTTP transport for the MCP server for proxy and long-lived deployments.
  • Checkpoint batch-save tool for multiple adddrawer calls plus diary entry in one round-trip.
  • Deletebysource bulk-cleanup tool for removing drawers and index entries by source file.

Improved

  • Multilingual embedding dependencies moved to core for out-of-the-box multilingual support.
  • Known-systems lexicon keeps multi-word product names atomic in entity detection.
  • embeddinggemma model now compatible with ChromaDB 1.5.x, fixing silently-failing semantic search.
  • Embedder no longer OOM-kills bulk re-embeds through optimized batch processing.
  • Backup retention prevents unbounded disk usage by cleaning up old backups from migrations and repairs.
  • Zero-config interpreter resolution for hooks using mempalace-mcp console-script shebang.
  • Added transcript parsers for Continue.dev sessions, Gemini CLI / AI Studio, and Pi agent JSONL.
  • Broader miner language coverage for C#/.NET, PHP, Swift, Kotlin, and Java with rootless subproject detection.

Fixed

  • Prevent silent data loss from drawerid hash collisions.
  • Importing mcpserver no longer recreates ~/.mempalace, respecting the privacy kill-switch.
  • diarywrite accepts a content alias and restores the missing-parameter diagnostic.
  • One-time wing-name migration normalizes legacy names to keep pre-upgrade palaces discoverable.
  • Fixed silently-no-op mining under uv tool by resolving Python interpreter from console-script shebang.

May 2026

2 releases: 3.3.4 → 3.3.5

May brought smarter setup with local language models, better cross-project connections, and critical fixes for palace storage and data integrity. Performance and reliability improvements made palaces smaller and more stable.

Highlights

  • Setup now uses your local language model for free, supporting Ollama, LM Studio, or llama.cpp with zero cloud dependency.
  • Cross-wing topic tunnels automatically connect ideas across projects when they share common themes.
  • A storage bug fix dramatically reduced palace disk size for long-running installations.
  • Setup and initial indexing are now combined into one step with automatic size estimates.
  • Palace repair tool can now recover from ChromaDB corruption issues.

New

  • Cross-wing topic tunnels link shared themes across different projects automatically.
  • mempalace repair --mode from-sqlite recovers palaces stuck on corrupt HNSW segments.
  • Setup offers optional immediate directory mining after onboarding with size preview.
  • GPU acceleration option for batch mining performance.
  • Windows UTF-8 support across stdio for proper rendering of non-Latin characters.

Improved

  • Setup uses local language models instead of cloud services when available.
  • Claude Code conversation recognition in background hooks prevents indexing as source code.
  • EntityRegistry.save() is now atomic and crash-safe during power loss or termination.
  • MCP tool search works reliably during the HNSW flush window after large mines.
  • Cross-process correctness on Windows and multi-tenant hosts with proper SQLite file locking.
  • Monorepo room detection routing fixed to prevent substring token false matches.
  • Compress command now handles palaces with over 32,000 drawers using paginated fetches.
  • Batch per-chunk upserts improve mining performance.

Fixed

  • Storage bug where active palaces bloated into multiple gigabytes over time.
  • Silent data loss in temporal graph triples now caught as write-time errors.
  • Claude Code conversations no longer misclassified as source code during indexing.
  • ChromaDB HNSW segment metadata staleness causing MCP search failures after bulk mining.
  • Windows file handle locking preventing palace reopening in multi-tenant environments.
  • Monorepo file path routing mismatching view directories with test files.

April 2026

6 releases: 3.0.0 → 3.3.3

April brought MemPalace to version 3.3.3 with major retrieval improvements, multi-language support, and extensive security hardening. The month focused on achieving state-of-the-art memory benchmarks while stabilizing the platform across different environments and use cases.

Highlights

  • Achieved 96.6% LongMemEval R@5 with zero API calls, the highest published score in its class.
  • Introduced Palace architecture with wings, rooms, halls, and other structures, gaining 34% retrieval improvement from structure alone.
  • Added multi-language support across 8 languages with script-aware processing for Arabic, Devanagari, Thai, and other complex scripts.
  • Launched BM25 hybrid search combining keyword and vector search for more reliable retrieval.
  • Released MCP server with 19 tools and automatic AAAK dialect teaching for any LLM.

New

  • Palace architecture organizing memory into wings, rooms, halls, tunnels, closets, and drawers.
  • AAAK compression providing 30x lossless shorthand that works with Claude, GPT, Gemini, Llama, and Mistral.
  • Knowledge graph storing temporal entity-relationship triples in local SQLite database.
  • Specialist agents feature giving each agent its own wing and diary.
  • Closet layer with searchable index pointing to drawers and support for optional LLM closets.
  • Cross-wing tunnels enabling explicit project-to-project links.
  • Diary ingest with day-based drawers that upsert as each day grows.
  • International language support in 13 total languages with locale-specific regex patterns.

Improved

  • R@1 retrieval improved from 0.42 to 0.58 (+38%) with regex closets; R@5 locked at 1.00.
  • Optional LLM closets push R@1 to 0.67 with bring-your-own endpoint capability.
  • MCP server integration with Claude Code via one-step install for Claude and Codex plugins.
  • Auto-save hooks for Claude Code with emergency pre-compaction saves.
  • BCP 47 language codes now case-insensitive, resolving PT-BR, zh-cn, and Pt-Br uniformly.
  • Knowledge graph thread safety with locks on close, query, relationship, timeline, and stats operations.
  • Entity registry now defaults to local-only mode without outbound Wikipedia calls.
  • MCP stdout redirection prevents library logging from corrupting JSON-RPC channel.

Fixed

  • Shell injection vulnerability in save hooks and input validation at all MCP entry points.
  • File size guards, symlink skipping, and SQLite connection leak.
  • MCP null args hang, cmdrepair infinite recursion, and 500 MB OOM guard.
  • Windows Unicode crash fixed with script-aware word boundaries for Devanagari, Arabic, Hebrew, Thai, Tamil, and Khmer.
  • HNSW and SQLite drift causing SIGSEGV crashes with quarantine and recovery mechanism.
  • Install regression where mempalace-mcp console script declaration drifted between pyproject.toml and plugin.json.

Summaries are written automatically from the official release notes (full changelog ↗); check the original notes before relying on a detail. MemPalace: pricing, features and alternatives · All changelogs