Lost in old chats.
Tuesday's half-finished migration is somewhere in a session you closed. So is the reason you stopped.
Hem · work history for coding agents
Pick up where you left off.
Hem keeps a structured history of the work you do with your coding agents (tasks, decisions, corrections, outcomes), so you can find your place again. Works with any agent that speaks MCP.
Works with Claude Code, Codex, and other MCP agents.
No transcripts. Nothing goes back into your agent unless you ask.
The gap
You did this already. Where?
Explained, again.
You've walked your agent through this codebase four times. Each new session starts from zero.
Decisions nobody wrote down.
You picked one approach over another for a reason. A week later neither you nor the agent can say why.
How it works
Three steps. The second one is just your normal day.
-
Connect over MCP.
Add Hem's MCP server to Claude Code, Codex, or any MCP-capable agent. It takes a few minutes.
# example: config shape only, final ID pending "hem": { "command": "npx", "args": ["[PACKAGE_ID]"] } -
Work as usual.
As you work, your agent records structured events to Hem: a task started, a decision made, a correction, how it ended. No transcripts.
-
Look back by episode.
Hem groups events into episodes (units of work) and patterns (what keeps coming up). Pull them when you want them.
Index
Exactly what Hem keeps.
Recorded
- Task started
- Task completed
- Task failed
- Task abandoned
- Decisions
- Corrections
- Tool summaries
- Interactions
Not recorded
- Raw transcripts
- Your full conversation content
- Automatic injection of history into your agent
- Background lookups of your past work
Each event is short, structured, and makes sense on its own. You can read one months later and still know what happened.
Episodes
A clean edge on every piece of work.
Events roll up into episodes, one per unit of work, so you see the shape of what happened without re-reading the whole session. Patterns show what keeps recurring across episodes.
Episodes across sessions
Structured work events hold across agents and sessions — tasks, decisions, corrections, and how work ended.
Decisions, written down
The choice that shaped the approach stays readable months later. No reconstructing it from a chat scroll.
Open threads, visible
Abandoned and unfinished work stays marked. Patterns surface what keeps coming up across episodes.
Control
Your history. Your call.
On request only.
Hem never pushes history into your agent. list_events runs only when you ask for it.
No transcripts.
Hem stores structured summaries of the work, not your conversations. So there's nothing hidden to resurface.
Append-only.
Nothing gets silently rewritten. What's recorded stays as it happened.
Erasable.
Want something gone? delete_task_usage removes it. You can erase all of it.
Sign-in by Clerk. Events stored in Postgres, tied to your account.
Across agents
One history, whichever agent you used today.
Claude Code on Monday, Codex on Tuesday. Hem connects over MCP and HTTP, so your work history isn't tied to one vendor, editor, or model. Switch agents and your place is still there.
- Claude Code
- Codex
- any MCP-capable agent
What's in Hem
Here's what works now.
“Pick up where you left off” is what we're building toward. Here's exactly where Hem is today.
Ships today
Shipped- Recording structured work events over MCP + HTTP
- Episodes and patterns
- Erasure, no transcripts, no auto-injection
- A personal usage dashboard (Clerk sign-in)
Building toward
Coming soon- A guided “pick up where you left off” flow In development
- A place-marker in your agent and on the web that shows where you stopped
- Shared history for small teams, later
We'll mark things as shipped here when they actually ship. Watch the changelog.
FAQ
Questions builders ask.
Isn't this just agent memory?
No. Memory tools pull facts out of your conversations to personalize the model, and they decide what gets fed back in. Hem doesn't personalize anything. It keeps a factual, append-only record of the work: what was attempted, what was decided, what got corrected, how it ended. Nothing goes back into your agent unless you ask. Memory shapes the model. Hem gives you the history.
So it's observability for agents?
No. Observability tools instrument execution so teams can debug runs. Hem records work at the level you think in (tasks, decisions, corrections) so you can pick prior work back up. The usage dashboard is where it starts, not the point.
Is it a vector database or RAG?
No. Hem stores structured events in Postgres. There are no embeddings of your conversations and no retrieval pipeline pushing context into prompts. You can read every record yourself.
Why not just scroll my chat history?
Chat history is a transcript, stuck in one tool, one session, one window, and nobody re-reads it. Hem gives you the shape of the work: the task, the decision that mattered, the correction you made, whether it shipped or got abandoned. It holds across agents and sessions, and it doesn't store the transcript at all.
What exactly do you store?
Structured work events: task started, completed, failed, or abandoned, plus decisions, corrections, tool summaries, and interactions. No raw transcripts. Details on the privacy page.
How do I delete my history?
Use delete_task_usage. You can erase any of it, or all of it.
Will Hem change what my agent does?
Not on its own. Hem never injects history automatically. Your agent sees past work only when you ask for it with list_events.
Which agents work with it?
Claude Code, Codex, and any agent that supports MCP. Setup steps are in the docs.
Does the “pick up where you left off” flow work today?
Not yet. Today Hem records your work and groups it into episodes and patterns you can look back on. The guided resume flow is what we're building next. See what ships today.
What does it cost?
Hem is free while we're early. We'll post pricing before anything changes. See pricing.
Is there a team version?
Not yet. Hem is built for individual builders right now.
Start here
Find your place again.
Connect Hem to your coding agent in a few minutes. Keep working the way you already do.