The hub stores raw transcripts, runs extraction and publishes snapshots. A node on each machine follows along, and keeps organising locally when the network is down. Agents read what matters into context through one fixed entry point. Many devices and many agents share one copy, and by default it only runs on a trusted private network. Raw transcripts are the only source of truth; memory is just an index pointing at them.
Figures measured from the repository on 2026-11-09.
You discussed the deploy script on 10-22 and settled on one path: build-release.sh[1]; x86 and ARM artifacts each ship with SHA256SUMS[2]; cold-start readings live in that docs note, not in the memory body[3]。
For an LLM, reading one organised sentence costs less than parsing scattered fragments; the more scattered the fragments, the more likely the model fills the gaps itself.
This is not a feature list but a set of trade-offs. Each one below maps to a constraint in the code; read these three first and every pipeline that follows has a reason.
Ordinals, timestamps, speakers, tool calls, mentions, dedupe and provenance are all computed in code. Only the semantic judgements — entities, facts, relations, mood, ambiguity — go to the model, one question at a time.
A memory row holds a conclusion, a pointer to the source, and provenance. Full conversations, tool output, code and logs stay in the raw layer; when needed, follow the pointer back instead of copying the text into memory.
Identity is a computed hash: move it to another machine, import it again, and it is still the same record; seq a display number is not identity. Wherever the hub moves, the number still points at the same thing.
One store, one pipeline. The difference is who plugs in and what has to be in front of it.
Three weeks of conversation, still connected the next day: who you are, how you felt, and how that thing turned out.
A project spanning three weeks, still readable by the next person: who changed what, why, and who owns the next step.
The hub itself needs no agent; each device runs a local node. A backup catches up entry by entry along the replication sequence and may only take over after verification; takeover rotates the credential, so anything the old hub still sends is rejected.
The diagram shows the main path only. Offline, half-partitioned, local distillation and merge-back are rehearsed as an interactive state machine on the Reliability page.On narrow screens, scroll sideways to see the full diagram.
Open the full topologyWrite and read are independent paths, separated only by a memory layer that can be rebuilt. Extraction failures, prompt revisions or a model swap never lose the original session.
Each agent's own converter writes sessions into one format; if ingest checks fail, the converter is fixed. The core accepts no private format.
Code locates first, then the LLM takes on semantic tasks. Every output carries only seq and a verbatim quote; no other fields are accepted.
Code checks everything that can be checked mechanically; then a separate model review flags problems without rewriting or re-extracting. Rejected drafts carry a reason.
The conclusion, handle and provenance enter the memory layer; the raw text stays in the transcript. Tests look at rows actually written — zero rows is an error.
By default you get a brief with handles — the marker at the start of each line — [1]; use it to fetch that memory. The turn pointer in the memory then takes recollect_turn you back to the raw words. Programs read json; reviewers use expand to see the evidence chain.
Each page stands alone, with previous and next at the foot. Seven include an interactive demo.
Each one answers a real trap we hit; nothing is complex for its own sake.
