PRIMA Memory

These notes cover every PRIMA Memory release up to 2.7.0, newest first. Each entry describes what changes for you. The notes leave out changes that only affect how GenAI Skills Academy builds and maintains the product.

Updates arrive through AI Business OS, and you restart Claude Code afterwards so the new version loads.

2.7.0

Released on 15 October 2026, this version searches your most recent work first and adds the session history in your project register as the last search step.

• Most recent first. Without a time filter, a search covers the last 7 days and widens to 30 days, then to all your conversations, only when nothing relevant turns up. A time filter you give is never widened, and every reply states the steps it searched.
• Session history as the last step. When your conversations have nothing, PRIMA Memory reads the session summaries in your project register for the current month, then the monthly archive, newest month first, as Searching Your History describes.

2.6.2

Released on 14 October 2026, this version fixes how PRIMA Memory chooses its source in a Claude Code install, tightens what it reads and logs, and corrects the documentation.

• Claude Code history read first. PRIMA Memory now reads Claude Code history whenever Claude Code has saved sessions for your workspace, and the installer pins Claude Code mode from the start. See How Indexing Works.
• No reading outside your workspace in Claude Code mode. When your workspace has no shared project log, PRIMA Memory no longer scans your other folders for one.
• Codex sessions scoped to your workspace. In Claude Code mode, PRIMA Memory now reads the Codex sessions started in your workspace, plus any that record no working directory.
• Time filters apply everywhere. A time_filter such as “this week” now applies to semantic results and to the deep extract search as well as to keyword matching.
• Deep search by project. A deep search with project_path now finds the matching session extracts.
• Start-up log no longer records secrets. The start-up diagnostics now log the names of CLAUDE_* and PRIMA_* settings and never their values, some of which can hold tokens.
• Launcher finds Node more reliably. start.sh uses the node already on your path first, and says which folders it checked when it finds none.
• Documentation corrected. The README now documents all seven tools and every setting, the background rebuild, host detection, installing from the email you receive when you buy, and every source PRIMA Memory reads. It also states the one network access, a download of the embedding model on first use. See Privacy and Your Data.
• Licence. The product now ships the GenAI Skills Academy Software Licence, with the MIT notice for the open-source project it builds on.

2.6.1

Released on 17 August 2026.

• Correct version number. The server now reports the version it ships as, read from its own package file.

2.6.0

Released on 16 August 2026, this version recalls sessions from more AI tools.

• VS Code Copilot Chat sessions. PRIMA Memory now reads Copilot Chat transcripts and labels them [VS Code] in results. Copilot records no workspace, so these sessions are searchable from any workspace on your account.
• Codex sessions. PRIMA Memory now reads Codex sessions and labels them [Codex].
• Clearer start-up reports. Each source now reports at start-up whether it is not installed, installed but empty, working, or failing, and the server counts lines and records it cannot read.

2.5.3

Released on 13 August 2026.

• No start-up hang on cloud-synced workspaces. The server could stop responding at start-up on a host that offers to tell the server its workspace location, when the workspace sat in a cloud-sync folder such as OneDrive. It now answers at once and finishes its checks in the background, and the folder scan stops after two seconds.
• Start-up failures always explain themselves. A crash during start-up now writes its cause to the log before the server exits.
• rebuild_index runs in the background. The call returns at once and the rebuild continues, newest sessions first. It pauses while you search and picks up where it stopped after a restart. See Rebuilding the Index.
• get_index_status. A new tool reports how many session files are in the semantic index and how far a rebuild has got.
• Coverage warnings. Semantic results now say when part of your history is not yet indexed, so a missing result is no longer mistaken for a genuine miss.

2.5.1

Released on 30 July 2026.

• Upgrade from 2.4 completes. The move to the 2.5 storage format failed on the databases it was meant to convert. It now deletes and rebuilds the index, which holds only data derived from your session files.
• Searches no longer time out on a large backlog. Indexing before a search now stops after 20 seconds and continues on later searches, newest sessions first.
• Problems are visible. If the semantic part of a search fails, the result says the results are keyword-only.

2.5.0

Released on 29 July 2026, this version removes every compiled component.

• One build for every platform. Storage now uses Node’s built-in SQLite module, and the embedding model runs on WebAssembly. Nothing needs rebuilding for a particular machine.
• Node 22.5 or later. PRIMA Memory now needs Node.js 22.5 or later.
• Search quality unchanged. The model and its results are the same, and the index rebuilds itself once after the upgrade.

2.4.2

Released on 13 June 2026.

• Calmer start-up messages. The start-up check now reports a missing local shared log as information and no longer suggests that recall is broken.

2.4.1

Released on 11 June 2026.

• Shared log entries reach semantic search. Entries from the shared project log now embed correctly, including projects with several entries.

2.4.0

Released on 11 June 2026.

• Semantic search finds genuine matches. A scoring fault dropped good matches from semantic results. The index now uses the correct measure and rebuilds itself on first load.
• Results keep their source. Semantic results now show whether they came from your own transcripts or the shared project log, and who wrote a shared entry.

2.3.3

Released on 11 June 2026.

• Shared log in semantic search. Shared project log entries now reach the semantic index as well as keyword search.

2.3.2

Released on 11 June 2026.

• Better workspace discovery. The server now asks the application that starts it for the open workspace, and also reads CLAUDE_PROJECT_DIR, to find the shared project log.
• Start-up diagnostics. The start-up log now shows how the server located your workspace.

2.3.1

Released on 11 June 2026, this version changed nothing in a Claude Code installation.

2.3.0

Released on 11 June 2026, this version adds the shared project log as a source.

• Shared project log. PRIMA Memory now reads the log at .claude/state/sessions/ in your workspace. In a workspace shared through a sync service, your colleagues’ entries become searchable too.
• Source labels. Every result now says whether it came from your own transcript or the shared log, with the author of a shared entry.
• Start-up self-check. The server checks each source at start-up and warns when one is missing or empty.

2.2.0

Released on 27 May 2026.

• Settings for locations. PRIMA_SESSIONS_DIR and PRIMA_DB_DIR let you choose where PRIMA Memory reads sessions and stores its index. See Configuration.
• Faster session lists. get_recent_sessions and get_session_details now open only the files they need.
• Works with /save and /resume. /resume in AI Business OS accepts a time such as 15:40 and matches it against your recent sessions.

1.0.0

Released on 17 February 2026, the first release gave Claude Code tools to search sessions, list recent sessions, read a session, trace a file’s history, summarise a period and rebuild the index.