The promise
A later agent should be able to find where the work and relationship were, then continue with care. That does not require pretending the base model has changed weights or secretly remembers the user.
AIppocampus keeps the source trail reachable, then gives agents language for what each recall result can safely support.
Action grammar
| Signal | How an agent should use it |
|---|---|
direction_only | Use as scent or background direction. Do not make factual claims from it. |
direction_with_ref | Use as a candidate route. Reopen source before claiming. |
reopenable_route | Follow the route or handle before broad manual search. |
bounded_evidence | Use within declared scope. Deepen for exact wording, sensitive facts, or wider claims. |
source_open | Reason from the open source while respecting scope and privacy limits. |
ignore_or_blocked | Do not let it shape the answer except to explain the boundary. |
Retrieval path
provider transcript
-> clean source
-> registry and local indexes
-> search, MCP, hooks, sync
-> source reopen when claims matterHigh scores mean "look here first." They do not mean "this is true." For strong answers, exact quotes, stale claims, or private context, source reopening owns the claim boundary.
Read next
The architecture page shows where this grammar fits in the system. The evidence page shows how claims are separated from not-yet-proven material.
Adapted from the public wiki pages Architecture and Concepts and Source and Retrieval Pipeline.