01
Your history is never rewritten.
The full thread lives in a local store on your machine and stays intact. No model ever replaces your record with its own summary of it. What you said in week one is still there, word for word, in week ten.
CONTEXT · NO COMPACTION
Your thread never forgets.No compaction — ever.
Agent threads that run for weeks without losing the plot. Effectively infinite context, local-first, with the models you already use — and day 30 remembers day 1, verbatim.
day 1
Going with Postgres advisory locks for checkout — remember that, it's load-bearing.
· · · 29 days and hundreds of turns later · · ·
day 30
Careful — this retry path conflicts with the advisory-lock decision from day 1. Refactoring around it.
THE COMPACTION TAX
You've seen the message.
Summarizing earlier conversation to save space…
Most agent harnesses hit a wall partway into a long thread: the history gets compacted into a summary, and the conversation quietly continues from the summary instead of from what actually happened. Then the bill comes due:
- The constraint you set in week one silently disappears from the conversation.
- The agent re-asks questions you answered a hundred turns ago.
- Work gets quietly redone because the record of doing it was summarized away.
- Long threads drift: each recap is a copy of a copy, and the details blur.
WHAT MANNA CODE PROMISES
Three things that are simply true of every thread.
02
The compaction moment doesn't exist.
There is no threshold where the thread gets folded into a recap and continues from the recap. The “summarizing earlier conversation…” message never comes, on day 1 or day 300.
03
Any moment comes back verbatim.
The decision from week one. The exact query from turn twelve. The error message that started it all. When the work needs a past detail, it comes back as it was written — not as a summary remembers it.
IT COMPOUNDS
A brain that grows with every thread.
The work your tokens pay for is kept. Nothing is compacted away, so the analysis from week two is still there in month six, and your agents and bots draw on it whenever the work needs it. A compacting agent keeps a summary instead. Whatever fell out of it gets worked out again, and you pay for those tokens twice.
Manna Code A compacting agent Work it pays to redoILLUSTRATIVE
WHAT YOUR AGENTS CAN DRAW ON
Every step is on record.
Each response keeps its steps: the tools it ran, the code, queries and commands, the answer, and where it saved what it made.
What it makes stays put.
Reports, charts and files land on your disk, and the thread knows where they are. Month six can open what week two built.
Bots know their threads.
A Manna Bot runs its threads and can read their history, so new work starts from what those threads already learned.
Notes that outlive a thread.
Decisions and conventions saved to project memory come along into every new thread in that folder.
THE DIFFERENCE IT MAKES
Ask about week one, on day 30.
Illustrative — the same question, answered from a compacted history versus an intact one.
Most coding agents
◐ Context low — 4% until auto-compact
✳ Compacting conversation history…
↓ the thread continues from the summary
“You went with… some row-locking approach, I believe? Worth re-checking before I touch it.”
Manna Code
● Day 30 — thread history intact
weeks of work — every turn still on record
↓ nothing was ever thrown away
“Postgres advisory locks — decided on day 1, and it's load-bearing. I'll work around it.”
TRUST
Boringly easy to verify.
The thread lives on your machine. Manna Code is local-first: your history is yours to keep, inspect, and back up — not a transcript in someone else's cloud.
The old way ships in Settings, switched off. Old-school compaction — fold the whole history into a summary and carry on — is available as a compatibility toggle. It is off by default. That's how confident we are you won't want it.
And you stay in control. A built-in context browser shows exactly what the model sees, live — pin what matters, drop what's stale. See all features.
HONEST LIMITS
What we are — and aren't — claiming.
- Model context windows are still finite — no tool changes that. What never happens here is your history being destroyed to cope with it.
- A month of work isn’t crammed wholesale into every model call. The record is always intact, and the agent draws on it as the work requires.
- This page makes capability claims only. Cost is a different question — measured numbers live on the benchmark page, and we don’t publish claims we haven’t measured.
- Works with the models you bring — frontier APIs or local models with small windows.
LOCAL-FIRST · BYOK · NO COMPACTION
Pick up on day 30 exactly where day 1 left off.
Put Manna Code on your machine, bring the model keys you already use, and let the thread run as long as the work does.
CAPABILITY PAGE · NO COST FIGURES HERE BY DESIGN · MEASURED COSTS LIVE AT /BENCHMARK