nodecode-agent-si
Agent SI: sourced memory that keeps the user's own words as evidence, and searches the agent sessions already on this machine. A Nodecode rewrite of joelbrilliant/hermes-agent-si by joelbrilliant.
Two stores, one face. The brain holds what is true now as claims that cite a source and a quote, behind an evidence gate: a quote must be at least five consecutive words the user actually typed in this conversation, and a second source with another value makes a fact disputed instead of overwriting it. The session index holds what happened before, read from the Claude Code, Codex and Grok session files already on this machine. Nothing here runs a model, needs a key or dials the network: the search is keyword ranking over both stores, merged into one list.
Get it
Tell the agent "add the agent-si cell", or run nodecode add nc://agent-si.
Use it
The model gets three tools:
| Tool | What it does |
|---|---|
brain_search {query, limit?, side?} |
One ranked list over the brain and the past sessions. Each hit says which side found it; side narrows to truth or history. |
brain_get {id, offset?} |
A fact key with every source and whether they disagree, a note with the source message, or a session's own text, paged. |
brain_remember {fact, quote, key?, value?, title?, kind?} |
Keep what the user said. The quote must be five consecutive words from one of their messages, and that message is kept as the source. With key and value it is a tracked fact; without them a sourced note. |
/agent-si shows what the brain holds, /agent-si QUERY searches it, and
/agent-si index re-reads the session folders. (help :agent-si) is the
manual page.
Before each turn the cell puts a compact recall block in front of the message
being answered — up to prefetch_hits hits, ids and snippets, framed as data
rather than instructions. Facts come from the tools, never from the model's
memory: if the brain has nothing, it says so, and a disputed fact is a
question for the person rather than a value to pick.
Where it keeps things, all of it local:
brain/facts.jsonl,brain/records.jsonl,brain/sources.jsonl: the claims, the notes and the user's messages they cite.brain/index/<harness>-<digest>.txt: one excerpt file per session, with its origin on the first line. The session files themselves are only read, and never edited.
The agent-si section of config.jsonc takes brain_dir (empty keeps the
brain under the Nodecode home), index_sessions to turn the session index
off, session_roots for other folders of session files, recall to turn the
per-turn block off, and prefetch_hits. Nothing is required.
What it does not do
The original vendored an engine of change packets, pending advice and its own
SQLite index over Hermes's state.db; this rewrite keeps the facts, the
evidence gate and the session search, and reads session files instead of
opening another app's database. The demo brain, the backup and restore hooks
and the background refresh thread are gone: the index is built at start and by
/agent-si index, which is one keyword scan and takes a moment.
Run .github/test-cell.sh tools/nodecode-agent-si from the cells repository.
The tests cover the lifecycle, a planted session file found and paged, the gate
refusing a quote the user never typed, a fact disputed by a second value, a
note that outlives a restart, and the recall block.
