The Pipeline
The pipeline runs in the background so your agent does not have to manage its own memory mid-conversation. Most of it happens automatically. You talk. It learns.
Stages
Section titled “Stages”| Stage | What it does |
|---|---|
| Scribe | Extracts structured deltas from conversation buffers. Captures evolving opinions, frustrations, and contradictions, not just hard facts. |
| Auditor | Mechanical triage. Detects stale nodes, contradictions, and noise candidates. |
| Librarian | Applies judgment-heavy graph updates with a prune-over-preserve philosophy. Regenerates context files. |
| Dreamer | Creates speculative cross-node associations via creative recombination. |
| Observer | Produces structured observations from conversation patterns. Writes to the shared nodes/ store. |
| Skillforge | Promotes frequently-accessed nodes into executable slash command skills. |
| Bootstrap | Auto-generates project docs (CLAUDE.md, AGENT.md) from mental model data. |
| Working update | Extracts key files from tool traces to prime the next session with what you actually edited. |
Runtime modes
Section titled “Runtime modes”The pipeline runs either in manual mode (just the tool, no daemon) for lightweight local testing, or Docker daemon mode (recommended) where the host agent stays on your machine and bounded workers run in a container against the mounted graph root.
Configure the runtime with the configure_runtime action or --docker at install time.
Reliability
Section titled “Reliability”If a worker times out (for example, a provider usage limit), the daemon retries on the configured fallback worker (fallbackProvider/fallbackModel). Daemon ticks are wrapped in try/catch, per-file I/O is guarded, and unknown job types throw explicit errors rather than failing silently.