How Templi learns your firm's way of working
Templi keeps two stores that outlive any one conversation: Knowledge, which is what is true about your firm, and Memory, which is what is true about one client. Both are read by the agent before it writes anything, and both grow as you work rather than through a setup project.
Why two stores
They answer different questions, and separating them is what stops your house style leaking into one client's file, or one client's circumstances leaking into every report.
Knowledge | Memory | |
|---|---|---|
Scope | The whole firm | One client |
Holds | Your terminology, house standards, how you write | Decisions, objections, goals, changes in circumstances |
Where you find it | Settings, then Knowledge | The Memory tab on a client |
Changes when | You tell Templi how the firm works | A fact find, meeting or document completes |
Knowledge overrides the agent's own defaults
An entry is not a hint. Where your knowledge disagrees with what the agent would otherwise write, your firm wins, and that includes wording it would normally consider standard.
A firm that writes its reports in Spanish, or calls the ongoing advice charge something specific, or orders its suitability sections a fixed way, gets that in every report once it is an entry. You are not re-explaining it in each chat.
Not everything is loaded every time
Each entry says when the agent should read it, because an agent carrying every entry on every run has less room for your client's actual circumstances.
Always load puts the entry in front of the agent in full, on every run. Suited to short, universal entries such as a charges glossary.
When relevant shows the agent the entry's name and description, and it opens the entry when the work calls for it. This is the right default.
Manual only hides the entry until someone names it.
An always-loaded entry also names which agents read it, so report wording can reach document generation without also reaching your meeting write-ups. A dot beside an entry in the tree means it is always loaded.
Keep always-loaded entries short and few. Long entries belong on "When relevant", where the agent opens them only for the work that needs them.
Memory builds itself as you work
Entries appear on a client's Memory tab as fact finds are completed, documents are finalised and meetings take place. The agent also writes an entry when it learns something durable in conversation: a decision, an objection, a goal, a vulnerability, or a preference for how the client wants to be dealt with.
A one-off instruction for the task in hand is not memory. Something is recorded when it should still matter the next time someone works with that client.
Templi asks when a new fact contradicts an old one
When what the agent has just learned conflicts with something already stored, it stops and asks you rather than choosing. It quotes both the stored wording and the new wording and names the entry, so you can see exactly what would change.
This matters most for firm knowledge, where a silent overwrite would change every future report. Your own standard is never replaced without you saying so.
What Templi will not record
The agent records what your firm has told it. It does not invent regulatory, tax or compliance guidance and store that as your firm's position.
Templi baseline entries are read-only. To replace one, create your own entry with the same name in the same collection, and yours is used instead.