What it is
AIppocampus keeps conversation source reachable across threads, projects, devices, and long-running context. It helps a later agent find prior evidence and continue carefully, without claiming the foreground model has innate memory.
It can support project work, but it is broader than a task memory system: old corrections, unfinished work, style preferences, recurring questions, reading notes, and relationship context can all matter when a new thread begins.
The six layers
| Layer | Job | Boundary |
|---|---|---|
| Conversation source | Preserve provider-owned transcripts. | Audit path, not the daily recall surface. |
| Clean source | Keep visible user turns and final answers. | Original wording, not summary memory. |
| Registry and indexes | Find threads, snippets, and source refs. | Search is navigation until source backs it. |
| Ambient hooks | Notice when old context may matter. | Quiet, opt-in, and safe to fail. |
| Semantic routes | Organize attention and propose routes. | Helpful direction, not fact. |
| Access surfaces | Expose CLI, MCP, sync, and package entrypoints. | Public contracts stay additive. |
What it is not
- Not proof that a model remembers you internally.
- Not a generic hosted RAG service.
- Not a simple vector database with friendlier wording.
- Not a cloud profile service.
- Not a replacement for reopening source when claims matter.
Source trail
Use the repository docs for implementation detail and contracts.
Adapted from the public wiki page Architecture and Concepts.