Long-term memory
for your agents.
Your agents forget everything between sessions, and their history is scattered across tools you can't search. Moraine keeps the complete trace of every session — exactly what happened, not a summary — and makes it searchable, so agents can recall everything they've ever done. It all stays on your machine.
uv tool install moraine-cli
the ridge of rock and sediment a glacier deposits as it moves and melts away.
Your agents are the glacier. Moraine is the record they leave behind.
Markdown memory forgets.
The traces are the truth.
Every other "agent memory" asks the model to write down what it thinks mattered. That push model is lossy by definition — it editorializes, compresses, and the instant the note is written the real detail is gone. Moraine never summarizes. It records the complete session trace and lets agents pull from it — you read what actually happened, not a story about it.
The agent writes notes.
↳ only what the agent chose to keep — the rest is gone.
A summary chosen in the moment: editorialized, partial, and stale the instant it's written. What it left out, you can never get back.
Moraine keeps the trace.
↳ every turn, tool call & token — recorded, queryable.
The raw record of what actually happened. Agents pull from ground truth — and you can always re-read the exact session, byte for byte.
One unified record of everything your agents did.
Conversation turns, tool calls, token counts and timestamps from every harness, in one consistent schema you can browse, query, and search.
For anyone who runs agents all day — across more than one tool.
Query the raw record
Turns, tool calls, tokens and timing land in one open schema. Point your own SQL, dashboards or notebooks at moraine.events — no lock-in, no black box.
Realtime ingest + backfill
Watches Codex, Claude Code, Cursor, Kimi, Hermes and Pi session files live, and indexes your entire existing history on first run.
Monitor UI
Browse sessions, inspect live analytics, and check ingest health at 127.0.0.1:8080 — token, turn and concurrency charts included.
Recall for your agents
Agents search their own past decisions, fixes and errors on demand — the whole record, exposed as tools they already know how to call.
Cross-provider search
One keyword search spans every harness at once. Find that decision, fix or error wherever it happened — whichever agent did the work.
Fully local by default
Runtime state lives under ~/.moraine. Nothing leaves your machine unless you point Moraine at remote infrastructure.
From scattered session files to one searchable history.
Six harnesses. One trace stack.
Moraine ships ingestion adapters for the agent tools you already use — every source runs live and backfills on first start, enabled by default.
Give every agent recall of every session.
Add a few lines to your harness instructions and your agents stop asking you what already happened. They search the record themselves — across providers, in realtime.
list_sessionsEnumerate recent and active sessions, any harness.search_sessionsKeyword search over the full trace history.openPull a specific session's turns, tools and context.
Watch your whole agent fleet, live.
Generation tokens per model, turns per model, concurrent sessions, and a searchable session list — updating in realtime at 127.0.0.1:8080.
Install and run in three commands.
Stand up your local trace stack, then point your agents at their own memory.
Then add the search guidance to your harness — your agents do the rest.