hermes-session-recall

Hermes Session Recall (session_search)

Recovering state from past conversations: where a task left off, what question was pending, what a session's final message said. Triggered by "prompt me again" after an approval timeout, "where did we leave off", or any recall request about earlier sessions.

Procedure

1. Discovery first, narrow: session_search(query=<distinctive keywords>, sort="newest", limit=3). Adaptive detail hydrates the top hit fully with an anchored window around the match — enough to answer most recall questions in one call. sort="newest" biases toward "where did we leave off".
2. To find how a session ENDED, phrase-match the tail: query for wording the final message would plausibly use ("waiting on you", "your call", "merge it when ready", "what remains"). The hydrated window lands ON the last message — which is the pending question/state you need.
3. Zoom in with the scroll shape: session_search(session_id=..., around_message_id=<match_message_id>, window=N). Keep windows small; add role_filter="user,assistant" to exclude tool noise.
4. Re-fire pending prompts: after an approval/clarify timeout, the user's "prompt me again" means re-present the SAME question — recover its text from the session tail (step 2), then re-issue it promptly. If recovery fails after one or two quick attempts, ask the user plainly what decision they're approving — never silently re-derive a different question.

When the prior conversation is NOT in this context

A follow-up that reads as a continuation ("write a doc about this", "same branch as before") often arrives in a NEW session created seconds earlier: none of the earlier conversation is in context, and the referring words ("this", "both", "the PR") are unresolvable from the message alone. Resolve the referent before acting on it.

1. Browse live activity, never infer it: session_search() with no args lists the most recently ACTIVE sessions (started_at / last_active + preview). The session whose last_active is minutes old is the conversation being continued — confirm from its preview, then read or scroll it.
2. Bound the browse with after="<date>") when several sessions compete. The bound applies to session START, so a long-running session that spans days falls outside a recent bound — drop the filter for those and rank by last_active instead.
3. Read the tail, then answer: scroll the found session near its end (a real id from that session; the LAST id of one window re-scrolls forward) and reconstruct what the last exchange actually asked, then act on that — not on the literal wording of the new message.
4. ~/.hermes/sessions/sessions.json is the gateway routing index (session key → active session id), not a session list: the key for the current thread points at the session created moments ago, while the conversation being continued sits under a different key. Use it to resolve a routing key only — never to find "the last conversation".