Personal Memory
Import local Codex memory into a private, searchable Intertool vault.
Personal memory is a private layer for durable context collected by your coding agents. It belongs to your user account, follows you across organisation memberships, and is not available to teammates.
This is separate from Team memory, where engineering conventions, decisions, warnings, and runbooks are reviewed and explicitly published for an organisation.
Automatic agent capture
Connected agents receive instructions to save useful verified learnings with capture_memory after substantive work, without a manual remember request. The tool accepts a repository, title, concise summary and evidence reference. It saves directly to My memory under the agent_capture source. Identical submissions return the existing memory; differently worded summaries are not semantically deduplicated.
At the start of later work, recall_memory searches your own captures for the requested repository. It returns at most five memories. Imported Codex documents remain dashboard only. Captures are private and never become team knowledge automatically. Use the existing draft and approval workflow to share a learning.
MCP supplies tools and connection instructions, not a stream of the whole conversation. Automatic capture depends on the agent following those instructions and the client allowing the call. For Codex and Claude Code, install the lifecycle adapter to run retrieval and completion checks through client hooks with a durable retry queue. There is no background transcript collector. Agents must respect requests not to save and report failed saves honestly.
Only concise durable learnings should be sent. The capture endpoint checks for known secret patterns, but that check cannot identify every sensitive value. Agents must exclude raw conversations, prompts, code files, terminal output, credentials and customer data.
Import local Codex memory
Codex memory is stored locally under ~/.codex/memories/. Create a personal Intertool API token, then run:
export INTERTOOL_API_TOKEN="itk_..."
pnpm memory:import:codexSet INTERTOOL_URL when the API is not running at http://127.0.0.1:3001. Set CODEX_MEMORY_DIR to import from a non-default memory directory.
The importer recursively sends Markdown documents only. Each document uses its relative path as a stable source key. Re-running the command creates new documents, updates changed documents, and leaves unchanged documents untouched.
For larger imports, My memory shows a compact distribution of documents by the first folder in that source key. This is a path-based collection view, not an inferred knowledge graph: Intertool does not draw semantic connections it has not stored. Small or single-collection vaults keep the simpler document, source, and import summary instead.
The current MVP is an on-demand import, not continuous sync. It does not import ChatGPT web memory: OpenAI documents ChatGPT memory and local Codex memory as separate systems, and Intertool does not receive a bulk ChatGPT-memory feed through MCP.
Privacy boundary
- Personal documents are always scoped by the authenticated user ID.
- Organisation members and administrators cannot list or open another user's personal documents.
- Personal documents are never returned by
/api/context/search,/api/context/get, or the team MCP context tools. The separaterecall_memorytool only returns the authenticated user’s own agent captures. - Importing never creates or publishes team memory.
A later extraction workflow can identify engineering-specific candidates and create reviewable Team memory drafts. Human approval remains required before those drafts become retrievable team context.