Memory for
AI agents
Agents forget everything between sessions. Thegram gives Claude Code, Codex, OpenCode, OMP and ChatGPT one memory they share with you and with each other: what was decided, what is still open and what happened, kept per project in a .gram file with every revision.
01 · What agents keep
Outcomes,
not transcripts
An agent records what will still matter next week, one fact per record, and leaves out chatter, guesses, raw logs and secrets. You read the same records in Thegram as a document, a graph, a timeline and the dashboard for the document's type: Project Tracker, Journal or Knowledge.
Decisions
What was chosen and why. When the team changes its mind, the old decision is marked superseded and the new one says what it replaces, so the reasoning is never lost.
Events
Shipped changes and incidents, with the time they happened and a link to the commit or pull request.
Tasks, risks and blockers
What is still open, with status and priority, until someone closes it. The next session starts from this list instead of from scratch.
Metrics
Numbers that were actually checked, with their unit and when they were observed.
Sources
The Slack message, email thread or document a fact came from, cited by the events and decisions that rely on it.
Links
Records point at each other: a change implements a decision, a blocker holds up a task, a new decision replaces an old one. A link can also point into another .gram document.
02 · How it works
Read first,
write only when authorized
Thegram gives agents a set of gram_* tools over the Model Context Protocol (MCP). Reading, searching, analysis and drafts are read-only; writing requires the person's request or explicit authorization in the conversation.
One memory per project
Each Git repository uses
@mounts/<vault-id>/<repository>.gram, and its worktrees share that document. Directory overrides in~/.config/thegram-memory/projects.jsonname an exact Cloud document.One record vocabulary
Every memory uses the same record types and statuses. The agent records decisions, tasks, risks and events with the status that fits, and the dashboard, search and every agent count them by one rule: a mitigated risk or a proposed decision is still open, a resolved risk or an accepted decision is done, and a superseded decision is closed.
A brief when the session starts
Before a coding agent answers for the first time, the plugin gives it the project's brief: its documents, the open risks and tasks, and the records that match your first prompt. Later prompts bring what other agents and people changed in the meantime. In ChatGPT, once you pick the plugin in a chat, it reads the memory before it answers.
Look up, then read only what is needed
The agent searches the memory and reads the few records or sections it needs, instead of loading a whole notes file into its context.
Every write is a revision
Updates compare the field values the agent read, or explicitly pin a revision. If a precondition changed, the write is refused instead of overwriting, and the agent reads again before deciding.
One reminder, not a nag
If a coding agent's turn edits files or makes a commit without recording anything, the plugin reminds it once to write down what changed.
Ask about the past
The agent can read the project as it was on any date, or see when and why a decision changed before it reopens the question.
03 · Coding agents
One plugin,
one API connection
Issue an API key in Thegram Settings → Agents, explicitly select its vaults and permissions, and connect the plugin. The same Cloud MCP works without a desktop app or DMG.
- Node.js 22 or later
- Claude Code · Codex · OpenCode · OMP
Install the plugin
Give your agent the canonical installation guide:
Install Thegram by following https://thegram.one/install/agents.mdConnect with your private key
In the installed plugin directory run
node dist/thegram-memory.mjs connect. Paste the key at the hidden terminal prompt, not into chat, and choose your Cloud vault. Rundoctorto verify the connection.Start a new session
What does Thegram remember about this project?Revoke the key in Settings to disable it. Local-only files are not uploaded automatically.
04 · In the cloud
ChatGPT
and Claude
ChatGPT and Claude connect to the same Thegram Cloud MCP with OAuth. Sign in and approve the intended vaults. Connected Google Drive accounts require separate authorization. Reading or connecting another application does not itself authorize copying its content into memory.
- A Thegram account
- ChatGPT or Claude
ChatGPT
Install the official Thegram plugin from ChatGPT. Follow its on-screen prompts.
Claude
Use Claude’s MCP connectors with OAuth and its published identity. Connect Thegram in Claude.
05 · The tools
What the
agent calls
The same gram_* tools on the Mac and in the cloud. Each answer is kept small, so an agent spends its context on your work rather than on its memory.
gram_projects·gram_briefWhich folders, vaults and Google Drive accounts the agent can reach, and one call to get oriented in a folder or a document.
gram_recall·gram_contextSearch by relevance with date filters, then read many records in one call.
gram_read·gram_get·gram_listOne section of a document, one record, or the records that match a type, status, tag or date.
gram_create·gram_patchCreate a document, or change one against the values the agent read.
gram_changes·gram_rangeWhat changed since a time or revision, and when and why one thing changed.
gram_state·gram_relationsA document as it was at a revision or a date, and the links around one record.
06 · Questions
Questions
Still unsure? Email support@thegram.one.
How is this different from CLAUDE.md or AGENTS.md?
Those files hold instructions you write for an agent. Thegram holds what happened: decisions with their status, open work, events with dates and sources, and every earlier version. The agents keep it up to date themselves, and the same memory works across Claude Code, Codex, OpenCode, OMP and ChatGPT.
Is it a vector database?
No. The .gram file is the memory itself, in plain text you can open. Search indexes are built from it and can be thrown away and rebuilt at any time.
Where is my memory stored?
In the selected Thegram Cloud vault, or an explicitly authorized connected Google Drive account. Coding-agent keys and OAuth grants limit access. Local-only folders are not exposed or silently migrated.
Can I see and fix what an agent wrote?
Yes. Open the same document in Thegram. History keeps every revision and you can edit or close records yourself.
Can several agents write at the same time?
Yes. Field preconditions protect head-mode updates; writers can also pin a revision. A stale precondition is rejected instead of silently overwriting another change.
What does it cost?
Thegram is free.
Give your agents
a memory
Connect ChatGPT or Claude with OAuth, or connect your coding agents with a scoped API key. No desktop app is required.
- Connection guideChatGPT, Claude and coding agents
- The .gram filewhat a memory looks like
- thegram-mcpCoding-agent API-key plugin
- thegram-pluginChatGPT and Claude OAuth workflow
- llms.txtwhere AI agents start