Agent Historian is a reusable architecture and integration project for cross-session agent continuity.
It separates three concerns that are often conflated:
Agent Bookkeeper
-> capture, canonical session evidence, archive indexing, search, and provenance
Optional learned-context workflow
-> deliberate, agent-authored, evidence-backed project briefings
The project is intentionally not a replacement for source control, repository documentation, or human judgement. Current code, repository instructions, and decision records remain authoritative over learned context.
Archive/search is the accepted first product boundary. The Hindsight transcript mining pilot established useful limits, but is not an adopted integration: transcript-only extraction cannot authoritatively model a mutable codebase. Any later learned-context workflow is separate from Bookkeeper and is fed by deliberate agent curation over current project state and source evidence.
- Preserve raw session evidence before deriving indexes or learned summaries.
- Keep archive/search and learned memory independently rebuildable.
- Make source revision and provenance visible to every derived result.
- Treat the current checkout, repository documentation, and decision records as authoritative; learned context is supporting, revision-scoped interpretation.
- Fail open: a memory-system outage must not prevent an agent from working.
- Keep transport, storage, and deployment configuration operator supplied.
The companion Agent Bookkeeper project is the concrete capture, archive, and search product.
Run:
rtk git diff --check