Skip to content
Agents tracked: 442 Downloads (7d): 202M down 19.3% GitHub stars: 10.4M Releases (7d): 693 Agent status: 1 with issuesSince last limit reset: Claude 28d · Codex 1d Updated Oct 21, 2026
AgentGid
✦ Ask

Claude Mem changelog: what's new each month

Every stable Claude Mem release summarised by month: the highlights, new features, improvements, fixes and anything you need to act on. 5 months covered; the current month updates daily.

October 2026 · so far

22 releases: 13.29.0 → 15.0.1

October brought major architectural improvements to Claude Mem's search and recall system, a new team collaboration feature, and a redesigned request dashboard. The month also saw significant performance optimizations and the introduction of version 15.0 with a new Smart-first recall workflow.

Action needed
  • Old memsearch guided/auto API and flashsearch calls must be replaced with the new Smart workflow or explicit dumbsearch.
  • Hooks in exec form require Claude Code 2.1.139 or newer.
  • search, timeline, and getobservations tools are superseded by memsearch but remain available for older clients.

Highlights

  • Redesigned requests dashboard with cards, lists, and agent views showing live status and readable summaries of work in progress.
  • New Smart recall workflow that progressively fetches memories through search, timeline context, and detailed observations.
  • TeamBrain integration allowing users to join teams, sync work state to the cloud, and see team briefs in SessionStart.
  • Major performance improvements including indexed newest memories, local-first context building, and background cloud sync.
  • Support for Pi and DSH IDEs through new install options and progressive recall tools.

New

  • workstatewrite and workstateread MCP tools for appending and reading named lists with task tracking.
  • Hook savers that write events to local files and return immediately without waiting for the worker.
  • Pre-cached SessionStart context in ~/.claude-mem/state/context-cache/ refreshed every 2 seconds.
  • Prompted observer requests now compatible with provider prompt caching for reusable context.
  • memsearch MCP tool with guided and automatic modes for bounded, progressive memory retrieval.
  • TeamBrain team joining and team briefs showing open collisions, plans, decisions, and delegated items.
  • Pi and DSH IDE support with automatic and manual recall capabilities.
  • PostCompact hook that marks compacted threads to prevent timeline duplication.

Improved

  • SessionStart now fetches newest memories from an indexed view instead of sorting all observations.
  • Project-alias lookups no longer load individual rows, reducing database queries.
  • Worker builds context from local database immediately while sync hub pull runs in background.
  • Cowork plugin skips cloud read when claude-mem is installed locally.
  • Observer conversation recycling behavior now handles context-overflow errors gracefully.
  • tree-sitter CLI is downloaded at startup, verified by SHA-256, and picked up by running sessions.
  • Plan mode now allows read-only tools to run without permission prompts.
  • Search results now preserve authored JSON, code examples, and file references when selected.

Fixed

  • Private content no longer leaves the hook in server runtime.
  • Database upgrade race condition between older and newer workers.
  • Chroma connection locks no longer released by newer connections during reconnect.
  • Windows home folders with non-ASCII characters now work correctly.
  • Hooks no longer report project tsconfig as errors in React Native and Expo projects.
  • Viewer settings changes are no longer undone by old settings tabs.

September 2026

30 releases: 13.22.0 → 13.28.0

September brought major stability improvements, a new cost reporting feature for tracking spending, and a redesigned cloud sync system. The month was dominated by bug fixes and infrastructure updates across dozens of patch releases.

Highlights

  • Cost reporting with expense breakdowns, behavior metrics, and multiple output formats for analyzing Claude spending by session and project.
  • Redesigned cloud sync with automatic hub URL migration, better error handling, and a new Postgres-backed sync API.
  • Provider resolution improvements that let new installs work without manual configuration while keeping existing setups intact.
  • Fixed a critical worker bundle version mismatch that caused continuous restart loops in previous releases.
  • Enhanced telemetry tracking of model type, billing plan, and usage limits across all sessions.

New

  • Observed model and billing type tracking in session telemetry for better usage analytics.
  • Usage limit window details including limit type, overage status, and reset timing in telemetry events.
  • Host observer provider for local inference without API keys, with support for Cursor and Grok Bot IDEs.
  • Cost analysis report generation with self-contained HTML and PDF outputs, CSV line items, and JSON evidence.
  • Behavior metrics tracking frustration episodes, human/bot classification, and tool error patterns across sessions.
  • Automatic migration of legacy cloud sync hub URLs on startup without requiring user reconnection.
  • Sign-in links deferred to end of non-interactive installs so agents can present them to users.
  • Session cost tracking for transcript tokens priced at OpenRouter rates, with separate observer costs.

Improved

  • IDE and provider information now correctly propagated to analytics dashboards.
  • Single transcript read for both model detection and rate limit information, reducing I/O overhead.
  • Rate limit detection refactored to handle actual SDK rate limit message format instead of legacy format.
  • Cloud sync now respects Retry-After headers on 429/503 responses instead of hammering endpoints.
  • Cloud sync checkpoint validation made more tolerant to handle hub migrations without wedging.
  • OAuth polling now respects server-provided expiry times, up to 30 minutes for browser sign-in.
  • Provider configuration validation ensures saved keys are usable before attempting to use them.
  • Comprehensive install summary clearly indicates whether OAuth login ran, was kept from previous config, or was not required.

Fixed

  • Worker bundle version mismatch no longer triggers continuous restart cycles; worker validates its own bundle version and exits cleanly.
  • Rate limit quota guards and telemetry events now correctly detect rejection events from the SDK.
  • Corpus type filters preserved through corpus build, fixing incorrect unfiltered search results.
  • Observation titles no longer double-wrapped when using label-wrapped format.
  • Stale device ID operations no longer block entire sync outbox; invalid operations are now quarantined.
  • Windows credential reading for OAuth token lookup now compiles correctly.

August 2026

17 releases: 13.13.0 → 13.21.2

August brought major improvements to Claude Mem's memory observation system, trial signup flow, and cross-platform stability. The month focused on hardening the observer model, streamlining Pro onboarding, and fixing critical issues on Windows.

Highlights

  • Interactive mode-creator workflow guides users through building custom Claude Mem modes tailored to their domain and note-taking needs.
  • Memory observer now runs server-side for Pro users, with compiled observations automatically injected into every new session.
  • Streamlined CMEM Pro trial funnel with device-code approval closes a security gap and now offers 30-day trials consistently.
  • Fixed observer role to explicitly document that it operates silently without contacting or influencing observed sessions.
  • Comprehensive Windows support fixes including proper path expansion, Git Bash detection, and process cleanup for dependencies like Chroma.

New

  • Interactive /mode-creator workflow with guided capture-type, tag design, and custom mode installation to user storage.
  • Optional tag-triggered Telegram notifications with guided bot configuration and verification.
  • openBrowser() best-effort mode prints URLs for copy-pasting on headless systems.
  • Device-code approval in installer (XXXX-XXXX format) for secure credential pairing in the browser.
  • Live model pricing fetched from API instead of hardcoded in installer.
  • Observer-health alerting warns at session start when observations stop flowing.
  • 7-day Pro trial promotion surfaced across session-start banner, per-message banner, welcome hint, installer, viewer header, and cursor hooks.
  • Quota breaker mechanism withholds requests during cooldown and admits exactly one probe to re-check, persisting across worker restarts.

Improved

  • Observer prompt now explicitly lists all observation types and correctly describes available options.
  • Corpus filters, OpenClaw feed, and weekly digests recognize new security alert and security note types.
  • Worker error handling now passes through gateway error envelopes with unified logging (one line per failure instead of five).
  • Session-start warnings show correct message, action link, and request ID for Pro users.
  • OpenRouter app entry renamed to Claude-Mem with centralized attribution headers to prevent drift between worker and server.
  • Chroma concurrent-write fix built into shipped bundles.
  • Credential redaction hardened to handle short and whitespace-bearing values, any Authorization scheme, Cookie headers, and full URI userinfo.
  • Overflow spool re-spools remainder instead of dropping oldest events.

Fixed

  • Chroma process-tree cleanup on Windows and POSIX properly kills uvx → uv → python → chroma-mcp chain during shutdown with PID-identity checks.
  • tree-sitter.exe resolution no longer silently returns nothing on Windows.
  • Windows-style home paths (~/\) now expand correctly; POSIX paths with backslashes left untouched.
  • Git Bash preflight detection fails loudly with clear message instead of cryptic hook crashes.
  • Log tailing survives rename-and-recreate rotation via file-identity tracking.
  • Page restart route now requires successor process (different PID) rather than accepting the corpse.

July 2026

11 releases: 13.9.2 → 13.12.4

July brought major infrastructure improvements to Claude Mem, including a complete redesign of cloud synchronization, full Antigravity CLI integration, and widespread stability fixes especially for Windows. The month focused on reliability and correctness across cloud sync, worker management, and hook lifecycle.

Highlights

  • New cloud sync architecture (Durable Object hub, deterministic schema migration, durability guarantees, WebSocket speed layer, and comprehensive guardrails) for reliable multi-device memory synchronization.
  • Full Antigravity CLI support at feature parity with Gemini CLI, including dedicated installation, hooks, and context injection.
  • Major Windows stability sweep eliminating shell escaping issues, spawn hazards, and path handling bugs across all execution paths.
  • Fixed critical worker restart loop triggered by plugin upgrades that could trap users in infinite startup failures.
  • Eliminated 331 error-handling anti-patterns and removed extensive dead code and unused dependencies.

New

  • Gemini CLI host integration with adapter, installer, IDE detection, and hooks.
  • GET /api/sync/status endpoint reporting pending sync counts, last flush time, and last error.
  • /cloud-sync skill for status checks, credential migration, and troubleshooting.
  • Durable Object sync hub (Cloudflare Worker) serving isolated HTTP push/pull lanes per user.
  • Advisory WebSocket speed layer for near-real-time sync nudges.
  • New Release Branches guide documenting main, core-dev, and community-edge branches.

Improved

  • CloudSync flusher now debounces write bursts over 1.5s, uses single-flight logic, and implements capped exponential backoff on failure.
  • Prompts now correctly join through sdksessions to push real memory session IDs, enabling cloud-side views like Summary/Prompt toggle and Replay.
  • Worker cache directories now ranked deterministically by version (never mtime) across all four resolver types, preventing orphan-stamped directories.
  • SQLite settings writes now atomic with busy timeouts to handle concurrent worker and hook access safely.
  • Windows command spawning centralized to eliminate shell:true footgun; hooks now emit Windows-executable commands.
  • Codex SessionStart hook no longer fails at startup when sessionid is missing or transcript path invalid.
  • Semantic search now preserves Chroma relevance ranking instead of silently reordering by recency.
  • Worker identity converged to one bundle across cache, marketplace, MCP, CLI, and restart launches to prevent version skew.

Fixed

  • Critical infinite worker restart loop triggered by plugin upgrades (13.12.1 recommended for all users).
  • Worker restart delegation bug where stale workers spawned their own successors using outdated code; hooks now SIGKILL stale workers and spawn fresh (13.12.2/13.12.3).
  • Zombie-held worker ports now detected correctly, ending infinite startup-failure loops.
  • Codex adapter dropped explicit empty-string additionalContext instead of preserving it in SessionStart payload.
  • IPv6 literals in health checks and display URLs now bracketed correctly for Worker host.
  • Race condition where session memory ID registration during prompt upload could leave prompt permanently mis-keyed in cloud.

June 2026

5 releases: 13.7.0 → 13.9.1

June brought major telemetry improvements, comprehensive code cleanup, and enhanced reliability across the SDK. The release cycle focused on instrumentation clarity, provider consolidation, and fixing edge cases in job handling and search.

Action needed
  • Node.js minimum version raised from >=20.0.0 to >=20.12.0.
  • CI docker e2e job script renamed to e2e:server:docker.
  • Fabrication tracking fields removed from telemetry.

Highlights

  • Per-session telemetry rollups now aggregate observations and costs at session end instead of every 5 minutes, with new fields for tracking context-cache value.
  • Provider infrastructure simplified with a new OpenAICompatibleProvider base class unifying Gemini and OpenRouter lifecycles.
  • Error tracking now reaches PostHog with consent-gating and rate-limiting, controllable via a new kill-switch.
  • Ponytail audit removed over 10,000 lines of dead code and consolidated duplicate infrastructure helpers.
  • SDK reference documentation published under a new dedicated section.

New

  • CLAUDEMEMTELEMETRYERRORS kill-switch to disable error capture independently of analytics.
  • observerturnrollup field that sums observations and types across compression turns in a session.
  • contextinjectedrollup field tracking total observations injected and tokens saved versus naive approach.
  • OpenAICompatibleProvider base class for unified provider session lifecycle management.
  • CMEM-SDK Reference documentation.
  • Codex startup context loading through MCP with HTTP fallback.

Improved

  • Unified instrumentation path that sends full-fidelity logs locally and scrubbed/rolled-up data to telemetry.
  • Provider refactor preserves per-provider behavior via abstract flags while reducing duplication.
  • Worker-restart hardening with single-spawn gate for improved stability.
  • Telemetry scrubber extended to document all new rollup fields, ensuring only counts and sums are captured.
  • Job handling in generate() now transitions to failed state immediately on provider crash or parse error instead of staying stuck.
  • Search query behavior clarified to report filter-only status correctly instead of falsely claiming degradation.
  • Dependency preflighting with graceful degradation when CLI dependencies are missing.
  • Chroma shutdown hardened with uvx prewarming.

Fixed

  • Node version floor corrected to >=20.12.0 to match stdlib util.parseEnv availability.
  • generate() no longer leaves jobs stuck in processing state after provider failures.
  • search() empty-query path now correctly reports filter-only results instead of false Chroma claims.
  • Stale fabrication tracking references removed after upstream audit.
  • Codex installer no longer spawns shell unnecessarily.
  • Session initialization stabilized after server rename.

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