Skip to content

Open source Grok Bot alternative, Bring your own agent and subscription: message Claude Code, Codex, Gemini CLI, Cursor, Copilot and 40+ AI agents as bots on your own computer, from iPhone, Mac, Windows, Linux or terminal. Agent-agnostic and free. Also a Muse and Dots alternative. Built in Rust.

docs: carry a history bookmark in the rollover handoff

leepokaidf8db0db9356
diff --git a/docs/design/bot-memory.md b/docs/design/bot-memory.md
index 07b471c0..8dfd0180 100644
--- a/docs/design/bot-memory.md
+++ b/docs/design/bot-memory.md
@@ -7,12 +7,14 @@ tuned against the [evaluation](#evaluation).
Built differently from the plan below, for now:
-- Search is a substring scan over a few hundred entries, not an FTS5 trigram index.
+- Search is a substring scan, not an FTS5 trigram index. Memory has a few hundred entries; the transcript scan
+ took 0.6 s over 180k messages of one bot, so the index waits until a bot's search passes a second.
- No revision history per entry; the dream log and the chat lines carry undo.
- Proposals for a bot that asks first live in their chat card, not in the store, so there is no proposed state.
- Health, religion and politics get no special notice yet; voice transcripts aren't marked, and group rooms
don't feed dreaming.
-- Entries show who wrote them, not a link to the message they were learned from (the ids are stored).
+- Entries show who wrote them, not a link to the message they were learned from. The bot can follow it:
+ `search_memory` returns each entry's source messages, and `search_history` reads back from a message id.
- Until step 3, a change the user makes still re-renders the opening (and reloads a Claude session), and the
notice carries the whole memory rather than a delta.
@@ -370,7 +372,9 @@ The keep-list also lives in the opening, so harness-initiated compactions see it
- the last compaction summary;
- the last 3–5 user turns verbatim;
-- the open plan.
+- the open plan;
+- a bookmark: the id of the old session's last message. The new session reads back from it with
+ `search_history`'s `from` instead of guessing words, and pages on with the `next` each result page returns.
The user can edit it as a draft before it is sent (Amp's handoff works this way).
diff --git a/docs/features/context-and-memory.md b/docs/features/context-and-memory.md
index b4c15d15..284f73a0 100644
--- a/docs/features/context-and-memory.md
+++ b/docs/features/context-and-memory.md
@@ -96,6 +96,8 @@ the `codync-mem` crate; `chat/memory` runs them in the host.
starts fresh, but its memory remains.
- Memory search is a substring scan, not semantic vectors. Voice transcripts aren't marked yet, so dreaming
treats them like typed text. Group rooms don't feed dreaming.
-- `search_history` covers the bot's own chat, not group chats it took part in.
+- `search_history` covers the bot's own chat, not group chats it took part in. It finds messages by words, or
+ reads back from a message id (`from`, then each page's `next`); `search_memory` gives each entry's source
+ messages, so a bot can check where a memory came from.
Reply threads have separate session/context lanes; see [groups and threads](groups-and-threads.md). Group and delegated requests don't feed dreaming.