Team rollout
Sharing reviewed agent memory across a team
FAVA Trails keeps agent memory in a Git-compatible repository that your team controls. An agent saves a thought as a draft, a configured reviewer decides whether it becomes an approved record, and other agents see approved records by default. This page covers the decisions a team makes when several people and agents share one trail. Installation, configuration and exact commands stay in the repository documents it links to.
Draft
An agent saves a finding as a draft. Other agents do not see it in default recall.
Review
A configured reviewer applies a rubric or a probability threshold to that one candidate, with limited context. Its verdict is process control. It is not independent verification of the facts.
Shared trail
Approved records live in a repository you control. Peers receive them when they sync, after the writer has published.
What a team gets from shared reviewed memory
Teams rarely work through one AI surface. Engineers may use Claude Code or Cursor, someone else may stay in browser chat, and each tool keeps its own history. Without a shared record, people copy findings between sessions by hand, and a wrong note that one agent saved is trusted by the next agent that reads it. FAVA Trails addresses that with a lifecycle for each remembered fact. Drafts stay out of default recall until a reviewer approves them, each approved record carries its author, scope and lineage, and a correction replaces an approved record only after the correction is itself approved, with the original kept in history.
Recall by scope.
Agents recall approved records for a scope such as company/product/research instead of depending on one machine or one chat history.
Drafts stay private to their author.
An agent's own drafts are visible to it only through authoring mode, on a server configured with that agent's identity.
Lineage for corrections.
Supersession links a correction to the record it replaces. It records that an approved replacement exists. It does not prove the replacement is true.
A history you can inspect.
Every saved thought is a commit in a Jujutsu repository, so the team can read who changed what and when with ordinary version-control tools.
Rollout decisions and their tradeoffs
Start with the smallest workflow where shared memory already costs your team time. The decisions below are the ones that change how the rollout behaves. The linked repository documents own the commands and the exact configuration.
-
Create one trail repository the team controls
The data lives in a Git repository separate from the FAVA Trails server, so the server can be upgraded or replaced without moving the memory. Use scopes that mirror how people look for things, for example
company/product/researchorclient-name/program/risk. Setup is in the setup instructions. -
Decide who writes, and give each writing process its own identity
Each MCP process has one identity, set by the operator through
FAVA_TRAILS_AGENT_ID. The operator chooses it as a stable role name, not a host fingerprint, so a new machine does not need a new one, and processes that reuse a name share its authoring view. A shared endpoint is one identity boundary, so agents that must not see each other's drafts need separate authenticated endpoints. A process left without an identity can read approved records but cannot write. KeepFAVA_TRAILS_OPERATOR=1for a separate operator endpoint and never enable it on an endpoint agents use. The governed recall guide describes the three views (governed, authoring and history) and the identity setup. -
Choose the reviewer, which is also a decision about where text goes
A hosted reviewer receives the candidate before it returns a verdict, so a rejection does not mean the text stayed on your machine. A reviewer on a local OpenAI-compatible endpoint can keep the text on infrastructure you control, but FAVA Trails classifies an endpoint as local only by its loopback address. A tunnel or port forward can still carry the text elsewhere, and a provider label is not a privacy guarantee. The built-in default when nothing is configured remains the hosted OpenRouter reviewer, and existing installations resolve as they did before.
In a release that includes the local-reviewer bootstrap,
fava-trails bootstrapcan write a local reviewer configuration for new installations, and the release history says which release that is. With no review options, or only some of them, it writes only the policy, and a fresh incomplete configuration fails closed instead of calling a hosted service. Fields already present in the per-machine configuration can complete or override that policy, however, and a missing flag does not disable a reviewer that machine already has. Bootstrap reports when the effective destination differs from the one you asked for, and it can exit nonzero after it has created the data repository. Runfava-trails doctorto read the effective destination and model before the first promotion. The local-reviewer bootstrap was recorded on a locally built 0.9.0rc1 candidate installed from a local wheel, which is not a published package. A default install from PyPI does not give you that build. Check what your installation does withfava-trails versionand the setup instructions for the release you installed. The recorded walkthrough shows a local reviewer rejecting two drafts before approving a third. -
Decide how approved records reach peers
Publishing and fetching are separate steps. The
synctool fetches and rebases what peers have published. It does not publish by itself in every configuration.push_strategy: manualis the bootstrap default, and the writer publishes explicitly. Withpush_strategy: immediate, the server pushes after a successful write operation, includingsync. Manual publishing trades convenience for a deliberate moment before anything leaves a machine. The repository documents own the exact commands; the setup instructions link to them, so they are not repeated here. -
Connect two agents or tools first, then widen
Pick the pair that causes the most copy-and-paste today, for example a research assistant and a coding agent. Write down the identity and scope each uses before adding a third. Upgrades are covered in runtime and upgrade, and exposing a trail to a hosted client is covered in secure MCP tunnel.
What the team should plan around
Review is advisory. The reviewer sees one candidate at a time with limited context, so a convincing false claim can be approved, and it does not compare the candidate with records already in the trail. Teams that need a person to decide can record explicit operator approval on an operator-controlled endpoint, which the trail records differently from a model verdict.
Recall is lexical. A query is lowercased, split on whitespace and matched as substrings with AND semantics, so paraphrases and synonyms miss. Agents should be told to query with the words that appear in the records. Governed recall has the exact rule.
A successful tool return means the write finished. An interrupted write or replacement can leave a recoverable working copy, and the governed recall guide owns that recovery. The runtime and upgrade guide covers which install is loaded and how to upgrade it.
When a simpler store is enough
If agents only need a private scratchpad that nobody else reads, a plain file or a database is simpler and FAVA Trails adds little. If your team already reviews every shared note through pull requests and the writers are comfortable with Git, a repository of Markdown files may already do the job. FAVA Trails fits when several agents or tools write shared notes, when you want an approval step before those notes become the default answer, and when you want the lineage of corrections kept.