nc://optchat

OptChat Memory: endless local memory for a session. A Nodecode rewrite of aska-digital/optchat-memory by aska-digital.

Every message of every turn -- the operator's words, the agent's replies, its tool calls and their results -- is appended to a local log under the home, one JSON line a message, and never edited or deleted. From that log grows a binary tree of one-line summaries: each message becomes a line, adjacent lines merge in pairs, and the older the messages the more a line covers. Every request carries the resulting view, a fixed-budget tiling of the whole log, oldest first, so a session opened a year later still reads what happened. The agent opens any line back down to the verbatim message with two tools.

Based on Victor Taelin's OptChat spec, as the original plugin is.

Get it

Tell the agent "add the optchat cell", or run nodecode add nc://optchat.

Set it up

Nothing is required and no key is needed: the cell writes under the home and summarizes with the auxiliary model Nodecode already has. Its section in config.jsonc (or /settings) is optional:

A section with "enabled": false turns the cell off, and the log stops growing.

Use it

• Every request carries an optchat block: a header line with the count of messages remembered, then <chat> with one line a stretch of the log, id+n|text. A short message is its own line, word for word; a long one is summarized in the background, in order; a line not summarized yet reads (not summarized yet: zoom it).
• optchat_zoom (id, n) opens a line into the two lines of n/2 messages it was made from; n = 1 gives the message in full.
• optchat_date (id) says when a message was written.
• /optchat prints how much is remembered, how much is summarized and how the view fits; /optchat zoom ID N and /optchat date ID open a line from the command line.
• In eval: (optchat:zoom 2184 8), (optchat:message-date 2184), (optchat:status).

The log is the truth; the tree is only how the past is read. A summary that could not be written (no model reachable) leaves its line saying so and is built on a later pass -- nothing in the log is lost meanwhile.

What it does not do

• It does not replace the session's context: recent turns stay in the conversation, and the view is the deep past beside them.
• It does not delete, edit or age anything out; the log grows for as long as the folder is used. Remove the folder to forget.
• One writer per home: the original locks the log with flock so two processes cannot share it; here one Nodecode process owns the home, and a second gateway on the same folder appends to the same files without a lock.

Gaps against the original

• The summary call is one nle:complete per node on the auxiliary model, as the original's call_llm(task="optchat_compact") is one call; the original appends a size-correction turn and retries up to five times when the reply overruns the budget, and this rewrite cuts the reply at the budget instead.
• Budgets are characters here, bytes there; a UTF-8 character can cost more than one of them.
• The original keys its state per Hermes profile through hermes_home_key; Nodecode has one home per process, so this cell has one log per store_dir.
• Messages are dated to the second, without the original's UTC offset suffix.

Test it

.github/test-cell.sh tools/nodecode-optchat

Ten tests, offline: the cell installs and stops, a turn's frames land in the log and render in the block, a restart reads the same folder, a short message is its own line with no model call, a long one goes through a stubbed nle:complete, zoom opens a built line, /optchat answers, and a bad setting refuses.