History
Where it started
No grace at all, then three days, now seven — each with the reason it changed. An agent that can see this knows the shape of the problem, not just the latest answer to it.
flmnt → teams
Your team doesn't disagree about the code. It disagrees about what you decided.
One person changes their mind on Tuesday. Someone else's agent ships the old plan on Thursday. Both did their job — the thing they were working from was different, and nobody could see that until review.
The team ledger
A decision isn't a point. It's a path, and it's still moving.
Decisions evolve. What you settled in March gets adjusted in September and revised again in November — not because the first answer was wrong, but because the team learned something. Which means the decision point itself moves.
Retrieval tools treat a decision as a fixed point to find: a vector in a space, matched by similarity. That works for facts. It fails for decisions, because it can return a point without telling you the point has since moved, or in which direction.
The ledger is append-only. Every conversation point is written with who said it, what it was, and when, and nothing is ever edited or removed. So a decision arrives as a path — where it started, where it stands, and where the team is heading.
the path of one decision
Decision · superseded
No grace period. At the cap, calls stop.
The simplest thing that could work, chosen when the product had twelve customers.
Decision · superseded
Three days of grace after the cap.
Support needed time to reach the customer before anything broke.
Conversation point
Support raises that three days still isn't enough.
Recorded as said, before anyone had decided anything. This is the part that normally evaporates.
Decision · current
Seven days of grace after the cap.
The reasoning travels with it. The two versions it replaced stay readable as history without being served to anything.
Artifacts
The design and the spec are brought in line.
Merged
Pull request #491 carries the whole path as evidence.
A reviewer, or an auditor a year later, can walk from the merged code back to the sentence someone said in support.
Position
Where it stands
Seven days, current since Nov 3. This is the only part a retrieval tool can give you, and only if the words happen to say so.
Trajectory
Where the team is heading
Two revisions, both moving the same way: more room for support to reach a customer before anything breaks. That direction is inferable from the path and from nothing else.
Why trajectory beats a vectorGive an agent the current decision and it can comply. Give it the path and it can reason: an agent proposing a one-day window isn't just contradicting the current answer, it's arguing against the direction the team has moved twice. It can say so, and it can be right. That is the difference between a tool that recalls and one that understands the position it's working from.
And what it isn'tNot a transcript archive, and not search over your chat history. Those tell you what was said. This tells you what is load-bearing, where it came from, and which way it's going.
For the person who decides what the team uses
Two questions you currently answer with a meeting.
When I change something, how long until the whole team is working from it?
Not "when did I tell everyone" — when did every agent on the team actually start working from the new version. On the grace change, that was under a day, and nobody was told.
decision recorded · nov 3, 09:31
before: a message in a channel, and no way to know who acted on it
Which of my team's agents is holding a version we abandoned?
Because the chain records reads as well as writes, the answer is a list rather than an investigation. Anything that hasn't read the current entry is still working from the one before it. Here is the queue change from Nov 6:
priya's agent · currentomar's agent · currentci worker · currentines's agent · last read nov 3
Ines's agent hasn't had a session since the change, so it is still holding the version before it. Nothing is broken yet — but the next task it picks up in that area would be built on the old answer, and now that's visible before it happens rather than at review.
before: nobody could ask this question at all
Governance · enterprise
A memory tool that can serve a replaced decision is a governance problem.
That is the version of this a CTO recognizes. In a regulated context the question isn't productivity — it's whether you can show what was true when, who acted on it, and that the record hasn't been rewritten since.
Harness engineering is the practice of building the structure around an agent: what it may read, what it may touch, and how the result is proved. flmnt is the governed layer of that harness — the one that decides what every agent is allowed to believe.
- The evidence chainappend-only · who, what, when
- The trail is the record itself, not a log bolted beside it. Entries are never edited or deleted; a change is a new link that supersedes an old one. The same structure that answers your agents answers your auditor, and neither gets a different story.
- Governed contextwhat an agent may believe
- Retired decisions are readable but never served. That is a governance control, not a retrieval preference: it puts a boundary on what any agent in your organization can act on, and the boundary is auditable.
- Access and scopeprojects, not people
- Decisions are scoped to projects. Contractors and partners see the projects they're on. Removing someone removes their access, not the reasoning they contributed — the chain was never in their editor.
- Data handlingresidency · retention · export
- Your chain is your organization's record. Residency and retention to your requirements, export at any time, and never used to train anything.
- Evidence on the pull requestCI worker
- The chain is posted onto the pull request, where reviewers and auditors already look. Pairs with the proving layer of the harness if you adopt it — the workflow →
- Licensingnegotiated
- Enterprise licenses are negotiated, including deployment shape, residency and support. Teams start per seat with no minimum — pricing →
Team licenses
Per seat, no minimum, and one chain from the first session.
Bring the tools your team already runs. Setup is one config entry each.