Skip to content
Agents tracked: 378 Downloads (7d): 244M down 5.7% GitHub stars: 8.4M VS Code installs: 151M Releases (7d): 407 Agent status: 1 with issues Updated Oct 10, 2026
AgentGid
✦ Ask

Codebase Memory Mcp changelog: what's new each month

Every stable Codebase Memory Mcp release summarised by month: the highlights, new features, improvements, fixes and anything you need to act on. 3 months covered; the current month updates daily.

September 2026

1 release: 0.11.0

September brought improved memory querying with exact counts and better visibility into partial results, plus performance fixes for large codebases and macOS systems.

Highlights

  • Graph queries now return exact totals instead of approximate counts, with clear indication when results are truncated.
  • Continuation cursors let you fetch remaining results from any large query result set.
  • CPU budgeting per file now covers the entire file processing, preventing performance issues with pathological files.
  • Fixed crash when resolving cross-file dependencies in large C/C++ codebases.

New

  • Graph rows and node identities now visible in query results.
  • Truncation state indicator shows when you are viewing a subset of results.
  • Continuation cursors for paginating through omitted query results.

Improved

  • Per-file CPU budget now covers the whole file instead of just parsing.
  • Locale-free formatting on macOS eliminates process locale lock overhead on every vsnprintf call.

Fixed

  • Crash in cross-file resolution on large C/C++ trees.
  • Performance issue with pathological files due to unbounded CPU usage.

August 2026

8 releases: 0.10.0 → 0.10.8

August brought major architectural improvements to Codebase Memory, including a new coordination daemon for multi-session awareness, refined call graph precision with evidence-backed edges, and significant performance gains across all supported languages. The month focused on stability and correctness, with numerous fixes to installer behavior, cross-platform compatibility, and graph analysis accuracy.

Action needed
  • CALLS edge meaning changed: now means exactly one thing (invocation only); use new CALLREFERENCE edge for value passing patterns.
  • UI/non-UI archive variant split removed: all platforms now ship single archive with embedded UI.
  • uninstall --help flag no longer performs actual uninstall (previously had side effect of removing binary).

Highlights

  • New CALLREFERENCE edge class distinguishes invoking callables from passing them as values, closing precision gaps in higher-order patterns and callbacks.
  • Coordination daemon enables multiple cbm sessions to share state and coordinate, ending the isolation of separate MCP servers and CLI calls.
  • Major performance improvements: Java indexing reduced from 477s to 86s, C# from 1211s to 450s, TypeScript from 37s to ~24s on large codebases.
  • Tree-format tool output with TOON-compact prefix-grouped rows and real detectchanges impact reporting reduces output tokens by ~99%.
  • Single unified archive per platform with embedded UI replaces the old UI/non-UI variant split; automatic configuration of 43 coding agent and client surfaces.

New

  • CALLREFERENCE edge class for evidence-backed call graphs distinguishing value passing from invocation.
  • Coordination daemon backend allowing sessions to share store and coordinate.
  • checkindexcoverage tool (15th cbm tool) to parse and report coverage gaps in the graph per file.
  • Honest coverage reporting showing what did not make it into the graph as first-class data.
  • CBMRUNTIMEDIR environment variable to relocate the rendezvous directory in shipped builds.
  • Context column in searchcode results for UTF-8 output paths.
  • Cross-language suffix matching for qualified-name lookups.
  • UI enablement toggle in config without manual JSON editing.

Improved

  • Windows and Linux memory allocator behavior fixed, preventing freed pages from remaining committed in long-lived daemons.
  • Tree-format tool output with TOON-compact formatting and cycle detection.
  • Deterministic edge survival when the same logical edge is produced twice with different attributes.
  • CONFIGURES candidate collection now orders by content key before applying caps, independent of worker merge order.
  • Refusal messages now name the directory that actually refused and provide validation details.
  • Refused agent-config writes report the target, whether it exists, and file type.
  • Python aliased from-imports now resolve to the real definition.
  • Cypher aggregates fixed to count correctly with all truncation mechanisms working properly.

Fixed

  • Windows memory ratchet allocator override now compiles and functions on MinGW.
  • update --ui command restored after v0.10.2 removed obsolete UI/standard chooser flags.
  • Generated Pi extension now exports correctly, fixing Pi extension loading entirely.
  • Codex cbm tools now visible when CBMCACHEDIR is set by properly inheriting environment variables.
  • Installed skill file YAML parsing fixed by quoting description containing colons.
  • searchcode no longer echoes negative limits as result counts.

July 2026

1 release: 0.9.0

July brought major platform expansion, significant performance gains, and improved reliability across indexing and memory management. The release focused on Windows support, extraction accuracy, and handling of large repositories.

Highlights

  • Full Windows support with non-ASCII path handling, native ARM64 builds, and UI stability fixes.
  • Indexing is now 61% faster with deterministic output and automatic recovery from problematic files.
  • Graph UI scales to 10 million nodes with preview, filtering, and GitHub integration.
  • Memory management improved for large repositories with accurate accounting and use-after-free fixes.
  • Call and import extraction accuracy improved across 8+ programming languages.

New

  • Native Windows ARM64 build with end-to-end non-ASCII path support.
  • Crash and hang supervisor that quarantines problematic files and continues indexing the rest.
  • Graph UI node budget configurable up to 10 million nodes.
  • Dead-code filtering in the graph visualization.
  • Node code preview directly in the graph UI.
  • GitHub deep-links from graph nodes.
  • URL routing for graph navigation.
  • Offline label support in graph UI.

Improved

  • Indexing wall-time reduced by approximately 61% with deterministic output.
  • Linux RSS memory accounting is now accurate.
  • Peak versus live memory reconciliation for better visibility.
  • Bounded parallel retention for machines with limited RAM.
  • Graph UI rendering with contrast controls for better visibility.
  • Call extraction correctness across C, C++, Python, Go, PHP, Kotlin, Java, TypeScript, JavaScript, and Rust.
  • Import extraction accuracy across multiple languages.
  • Route extraction correctness improvements.

Fixed

  • UI hang on Windows resolved.
  • Use-after-free bug in cross-thread slab memory frees.
  • Installer and runtime reliability issues on Windows.
  • Call extraction errors in multiple programming languages.
  • Import resolution bugs across language implementations.

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